Awesome AI Models Matrix 🧠

Research-based list of AI models, development tools, and automation resources. Use it to compare releases, pricing, benchmarks, and deployment options from official sources.

🚀 Quick Start

Looking for the best AI models? Check out our Top Models by Category table below for quick recommendations. For detailed analysis, explore the comprehensive sections on frontier models, open-source alternatives, and development tools.

🔥 Highlights

138+ AI models from major providers (OpenAI, Anthropic, Google, Meta, xAI, Microsoft, and others)
Latest pricing and performance benchmarks (latest data through 2026-07-19 16:54 UTC)
GPT-5.6 reached general availability on 2026-07-09 across ChatGPT, Codex, and the OpenAI API; Sol, Terra, and Luna remain the three-tier family.
Grok-4.5 (xAI, 2026-07-08): new frontier multimodal LLM with a 500K context window, text+image input, function calling, structured outputs, and reasoning; API pricing $2.00 / $0.50 cached / $6.00 per 1M tokens. Now the xAI flagship route over Grok 4.3. Source: 🔗
Kimi K3 (Moonshot AI, 2026-07-16): new 2.8T-parameter open-weight frontier model with 1M context, native vision, $3.00 / $15.00 API pricing, and 90% cache-hit discount; open weights promised by 2026-07-27 under a modified MIT license. Source: 🔗
Meta Muse Spark 1.1 (Meta, 2026-07-09): new closed-weights frontier multimodal reasoning model for agentic tool and computer use, 1M context, available via Meta Model API in public preview for US developers. Source: 🔗
Inkling (Thinking Machines Lab, 2026-07-15): new 975B-parameter open-weight model with 1M context, native multimodal (text/image/audio), Apache 2.0 license, $1.87/$4.68 API pricing. Source: 🔗
Qwen 3.8 (Alibaba, 2026-07-19): new 2.4T-parameter open-weight model in preview on Token Plan/Qoder/QoderWork; claims to be "second only to Fable 5" on coding benchmarks. Source: 🔗
Codex CLI 0.144.5 (2026-07-16) improved dangerous-command detection and clearer rejection reasons; 0.144.4 (2026-07-14) had no user-facing changes; 0.144.3 (2026-07-13) was version-only.
Gemma 4 models were released through the Gemini API and AI Studio on 2026-07-06, including Gemma 4 31B and Gemma 4 26B-A4B.
Claude Sonnet 5 launched 2026-06-30 00:00 UTC with 1M context, 128K output, $2.00 / $10.00 introductory pricing through 2026-08-31 00:00 UTC, and 92.4% SWE-bench Verified
Claude Fable 5 and Mythos 5 access restored 2026-07-01 00:00 UTC after export controls were lifted; Fable 5 globally available while Mythos 5 remains restricted to Project Glasswing partners
Cartesia Sonic-3.5 and Ink-2 are now official production speech models for low-latency voice agents
MiniMax Speech 2.8 official page verified: native sound tags, high-fidelity cloning, and studio-grade multilingual TTS
GPT-Live (OpenAI, 2026-07-08): full-duplex voice model family powering ChatGPT Voice; GPT-Live-1 default for paid tiers, GPT-Live-1 mini for Free; no API yet (soon); GPT-5.5 backend delegates complex work. GPT-Realtime 2.1 / 2.1-mini shipped 2026-07-06 in the Realtime API with function calling, MCP, and SIP support
Cursor 3.11, Kiro 1.0.138, and Codex CLI 0.144.5 refreshed from official release notes
Codex CLI 0.142.5 patched WebSocket trace logging, and GitHub Copilot added enterprise agent session streaming plus GITHUB_TOKEN support for Copilot CLI in GitHub Actions
Browser Use CLI 3.0, Browser Harness, and Vercel agent-browser refreshed in Browser Automation from official project docs and releases
Vercel AI Gateway routing rules, eve Agent Runs in Vercel MCP/CLI, and Vercel Sandbox FUSE support added from official July 2026 changelogs
Cloudflare Browser Run /json Quick Action, Jina v5 Omni embeddings, and official MCP Registry references refreshed from official docs
MiniMax M2.7 self-evolving LLM with SWE-Pro 56.22%
MAI-1, MAI-1 Plan, and MAI-Frontier Tunings launched July 2026 with Microsoft Frontier Tuning framework reaching GA status
Cheapest verified direct APIs now tracked from official pricing pages: Groq Llama 3.1 8B Instant [$0.05/$0.08], Gemini 2.5 Flash-Lite [$0.10/$0.40], DeepSeek-V4-Flash [$0.14/$0.28], and OpenAI GPT-5.4 nano [$0.20/$1.25]
Google AI Plus price drop: $7.99 → $4.99/month with storage doubled to 400GB (effective June 8, 2026)
Self-hosting guides for open-source models
Development tools for AI application building

📊 Top Models by Category

Category	#1	#2	#3
Coding	Claude Opus 4.8	GPT-5.5	Claude Sonnet 5
Reasoning	Gemini 3.1 Pro	GPT-5.5 Pro	Qwen3.7-Max
Open Source	Kimi K3	GLM-5.2	MiniMax M3
Cost Efficiency	Llama 3.1 8B Instant	Gemini 2.5 Flash-Lite	DeepSeek-V4-Flash
Free & Budget	Gemini 2.5 Flash-Lite	Command A+	Mistral Small 4
Agentic Performance	Claude Opus 4.8	GPT-5.5	Gemini 3.1 Pro
Context Window	Gemini 3 Flash (10M)	Llama 4 Scout (10M)	Claude Opus 4.8 (1M)

📚 Documentation

Models 🧠 - Comprehensive model comparisons and specifications
Development Tools 🛠️ - IDEs, CLI tools, and API providers
Automation 🤖 - Browser automation and agent platforms
Guides 📚 - Getting started and selection guides
AI Infrastructure 🏗️ - Models, frameworks, and deployment options

📋 At a Glance

Document Version: 3.50
Last Updated: 2026-07-19 16:54 UTC
Repository: github.com/ReadyPixels/AI_Models_Matrix
License: CC BY-NC 4.0

🔍 Search & Navigation

Ctrl+F - Use your browser's search to find specific models or topics
Jump to sections using the table of contents above
Model search - Use your browser's search to find specific models (e.g., "GPT", "Claude", "Gemini")

Models 🧠

Comprehensive documentation of Large Language Models (LLMs), Small Language Models (SLMs), and specialized AI models available today.

Frontier Models 🚀

State-of-the-art proprietary AI models with cutting-edge capabilities from leading AI labs.

Model	Company	Context	GPQA Diamond	Arena Elo	SWE-bench Verified	AIME 2025	Pricing	Verified
Claude Opus 4.8	Anthropic	1M	93.6%	1506	88.6%	—	$5.00 / $25.00	2026-05-28
Claude Sonnet 5 ⭐	Anthropic	1M	—	—	92.4%	—	$2.00 / $10.00 (intro)	2026-06-30 00:00 UTC ⭐
Claude Opus 4.7	Anthropic	1M	94.2%	1505	87.6%	~95%	$5.00 / $25.00	2026-04-26
GPT-5.5	OpenAI	1M	93.6%	1506	88.5%	99.9%	$5.00 / $30.00	2026-04-26
GPT-5.5 Instant	OpenAI	1M	91.0%	—	88.7%	—	$0.20 / $1.00	2026-05-05
GPT-5.5 Pro	OpenAI	1.05M	94.2%	1551	92.3%	100%	$30.00 / $180.00	2026-05-08
Gemini 3.1 Pro	Google	1M	94.3%	1505	80.6%	100%	$2.00 / $12.00	2026-04-26
Claude Fable 5 ⚠️	Anthropic	1M	94.5%	1510	95.0%	—	$10.00 / $50.00	2026-06-09 ⭐
Claude Mythos 5 ⚠️	Anthropic	1M	~94.1%	—	95.5%	—	$10.00 / $50.00	2026-06-09 ⭐
NVIDIA Nemotron 3 Ultra	NVIDIA	1M	87.0%	—	—	—	Free (NVIDIA NIM)	2026-06-05
Kimi K3 ⭐	Moonshot AI	1M	93.5%	—	88.3% (Terminal-Bench 2.1)	—	$3.00 / $15.00	2026-07-16 ⭐
Meta Muse Spark 1.1 ⭐	Meta	1M	—	—	61.5% (SWE-Bench Pro)	—	API-only (Meta Model API)	2026-07-09 ⭐
Inkling ⭐	Thinking Machines Lab	1M	—	—	—	—	$1.87 / $4.68	2026-07-15 ⭐
Qwen 3.8 ⭐	Alibaba	1M	—	—	—	—	API-only (preview)	2026-07-19 ⭐
Kimi K2.7 Code ⭐	Moonshot AI	262K	65.8%	—	78.2%	91.5%	$0.95 / $4.00	2026-06-12 ⭐
GLM-5.2 ⭐	Zhipu AI	1M	~91.2%	—	—	99.2% (AIME 2026)	$1.40 / $4.40	2026-06-13 ⭐
MAI-Thinking-1 ⭐	Microsoft	256K	84.2%	—	—	97%	TBD (Private Preview)	2026-06-02 ⭐
MAI-Code-1-Flash ⭐	Microsoft	—	—	—	—	—	Free (Copilot)	2026-06-02 ⭐
MiniMax M3 ⭐	MiniMax	1M	—	—	80.5%	—	$0.30 / $1.20	2026-06-01 ⭐
MiniMax M2.7 ⭐	MiniMax	205K	—	—	56.22% (SWE-Pro)	—	$0.30 / $1.20	2026-03-18 ⭐
Qwen3.7-Plus ⭐	Alibaba	1M	—	—	—	—	$0.40 / $2.40	2026-06-03 ⭐
Step 3.7 Flash ⭐	StepFun	1M	—	—	—	—	Free (limited)	2026-05-28 ⭐
GPT-5.6 Sol ⭐	OpenAI	1M+	—	—	—	—	$5.00 / $30.00	2026-07-09 ⭐
GPT-5.6 Terra ⭐	OpenAI	1M+	—	—	—	—	$2.50 / $15.00	2026-07-09 ⭐
GPT-5.6 Luna ⭐	OpenAI	1M+	—	—	—	—	$1.00 / $6.00	2026-07-09 ⭐
Grok-4.5 ⭐	xAI	500K	—	—	—	—	$2.00 / $6.00	2026-07-08 ⭐

⭐ Kimi K3 (Moonshot AI, 2026-07-16): New 2.8T-parameter open-weight frontier model, 1M context, native vision, $3.00/$15.00 per 1M tokens with 90% cache-hit discount ($0.30 cached input). Open weights promised by 2026-07-27 under a modified MIT license. Source: 🔗

⭐ Meta Muse Spark 1.1 (Meta, 2026-07-09): New closed-weights frontier multimodal reasoning model built for agentic tool and computer use, 1M context, native multimodal perception, available via Meta Model API in public preview for US developers. Source: 🔗

⭐ Claude Fable 5 and Mythos 5 redeployed: Anthropic says export controls were lifted on 2026-06-30 00:00 UTC, and access to Fable 5 and Mythos 5 was restored on 2026-07-01 00:00 UTC. Fable 5 is available globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with cloud partner re-enablement in progress. Mythos 5 access is restored for approved US organizations and remains limited to Project Glasswing partners. Fable 5 has a 30-day data retention requirement, uses a different tokenizer (+30% token overhead vs Opus), and is unavailable for Zero Data Retention (ZDR) agreements. Source: 🔗

⭐ GPT-5.6 Sol/Terra/Luna (OpenAI, 2026-07-09): General availability release following the limited preview. Three-tier family: Sol ($5.00/$30.00), Terra ($2.50/$15.00), Luna ($1.00/$6.00) per 1M tokens. Source: 🔗

⭐ Grok-4.5 (xAI, 2026-07-08): New frontier multimodal LLM — the xAI flagship route over Grok 4.3. 500K context window, text+image input and text output, with function calling, structured outputs, and reasoning. API pricing $2.00 input / $0.50 cached input / $6.00 output per 1M tokens. Released 2026-07-08; available in Palantir AIP from 2026-07-14. Source: 🔗

⚠️ GPT-5.5 Instant pricing correction: The API pricing for GPT-5.5 Instant is $0.20/$1.00 per 1M tokens (input/output), not $5.00/$30.00 which is the standard GPT-5.5 rate. GPT-5.5 Instant replaced GPT-5.3 Instant as the free-tier ChatGPT default on May 5, 2026.

Model Specifications 📋

Detailed technical specifications, pricing, and capabilities for frontier models and notable API models. Data as of 2026-07-19 16:54 UTC.

Output Token Limits

Maximum output tokens per single API request.

Model	Max Output	Context Window	Notes
Claude Opus 4.6	128K (300K via beta)	1M	Extended output via `output-128k-2025-02-19` beta header
Claude Opus 4.7	128K (300K via beta)	1M	Extended output via `output-128k-2025-02-19` beta header
Claude Sonnet 4.6	64K	1M	—
Claude Sonnet 5 ⭐	128K (300K via beta)	1M	Released 2026-06-30 00:00 UTC; adaptive thinking on by default; batch API supports 300K with `output-300k-2026-03-24` beta header
Claude Sonnet 4.5	64K	200K	—
GPT-5.4	128K	1.05M	—
GPT-5.4 mini	128K	400K	—
GPT-5.4 nano	128K	400K	—
GPT-5.3-Codex	128K	400K	—
GPT-5.5 Instant	128K	1M	Default ChatGPT model since 2026-05-05
Gemini 3.1 Pro	64K	1M	—
Gemini 3.5 Flash	64K	1M	Released 2026-05-19 at Google I/O 2026; 4x faster than 3.1 Pro
Gemini 3 Pro	64K	2M	—
Gemini 3 Flash	64K	1M	—
Gemini 3.1 Flash-Lite	32K	1M	—
DeepSeek-V4-Flash	384K	1M	API model ID `deepseek-v4-flash`
DeepSeek-V4-Pro	384K	1M	API model ID `deepseek-v4-pro`; 75% discount made permanent 2026-05-31; standard price $0.435 / $0.87
DeepSeek-V3.2	8K / 64K (reasoner)	128K	Reasoner mode unlocks 64K output
Qwen3.5-Max	65K	1M	—
Qwen3.7-Max	65K	1M	Extended thinking by default; proprietary, API-only
GLM-5	128K	200K	—
GLM-5.1	131K	200K	—
MiniMax M3	512K	1M (guaranteed 512K min)	Native multimodal (text/image/video); open weights pending ~2026-06-11
MiniMax M2.7	262K	205K	Self-evolution capabilities; Agent Teams native; SWE-Pro 56.22%
MiniMax-M2.5	131K	1M	—
Kimi K2.6	262K	262K	Max output = 262,144 tokens (MIT open-weight)
Kimi K2.7 Code	262K	262K	Coding-focused variant of K2.6; 30% fewer thinking tokens; Modified MIT
Kimi K3 ⭐	128K	1M	2.8T MoE (16 of 896 experts active); open weights promised by 2026-07-27; $3.00 / $15.00
Meta Muse Spark 1.1 ⭐	—	1M	Closed-weights frontier multimodal reasoning model; agentic tool and computer use; API-only (Meta Model API)
Inkling ⭐	—	1M	975B/41B active MoE; Apache 2.0; multimodal (text/image/audio); released 2026-07-15
Qwen 3.8 ⭐	—	1M	2.4T parameters; open-weight promised; preview on Token Plan/Qoder/QoderWork; released 2026-07-19
Step-3.5-Flash	66K	256K	—
Grok 4	—	256K	Not publicly specified
Grok 4 Fast	30K	128K	Now aliased to Grok 4.3; $0.20 / $0.50
Grok 4.20	30K	2M	All variants: $1.25 / $2.50
Grok 4.3	30K	1M	Always-on reasoning; released 2026-05-01
Grok-4.5 ⭐	—	500K	Multimodal (text+image); function calling, structured outputs, reasoning; $2.00 / $6.00; released 2026-07-08
Mistral Small 4	8K	256K	119B MoE; multimodal (text+image); $0.15 / $0.60
Mistral Medium 3.5	8K	256K	Released 2026-04-29; 128B dense; $1.50/$7.50
Mistral Large 3	8K	256K	—
Llama 4 Scout	16K	10M	—
Llama 4 Maverick	16K	1M	—
Claude Fable 5	128K	1M	Mythos-class; $10.00 / $50.00; released 2026-06-09
Claude Mythos 5	128K	1M	Mythos-class (restricted); $10.00 / $50.00; released 2026-06-09
Kimi K3 ⭐	—	1M	2.8T MoE; open weights promised by 2026-07-27; $3.00 / $15.00; released 2026-07-16
Meta Muse Spark 1.1 ⭐	—	1M	Closed-weights; agentic tool and computer use; API-only; released 2026-07-09
GPT-4.1 ⭐	128K	1M	Knowledge cutoff June 2024; $2.00 / $8.00; retired from ChatGPT Feb 13, 2026; API still available; released 2026-04
GPT-4.1 mini ⭐	128K	1M	Knowledge cutoff June 2024; $0.40 / $1.60; retired from ChatGPT Feb 13, 2026; API still available; released 2026-04
GPT-4.1 nano ⭐	128K	1M	Knowledge cutoff June 2024; $0.10 / $0.40; retired from ChatGPT Feb 13, 2026; API still available; released 2026-04
GPT-5.6 Sol ⭐	128K	1M+	GA model; $5.00 / $30.00; released 2026-07-09
GPT-5.6 Terra ⭐	128K	1M+	GA model; $2.50 / $15.00; released 2026-07-09
GPT-5.6 Luna ⭐	128K	1M+	GA model; $1.00 / $6.00; released 2026-07-09
Gemini 3.1 Flash Image ⭐	—	131K	Image + text + audio understanding; generation; $0.50/1M text+image input tokens; output images priced by resolution ($0.045/512px, $0.067/1Kpx, $0.101/2Kpx); GA June 2026

Cached & Batch Pricing

Discounted pricing tiers for high-volume usage. All prices in USD per million tokens.

Model	Standard Input	Cached Input	Batch Discount	Notes
Claude Opus 4.7	$5.00	$0.50 (hit) / $6.25 (5m write)	50% off	Batch: $2.50 in / $12.50 out
Claude Opus 4.8	$5.00	$0.50 (hit) / $6.25 (5m write)	50% off	Same pricing as 4.7; fast mode $10/$50
Claude Opus 4.6	$5.00	$0.50 (hit) / $6.25 (5m write)	50% off	Batch: $2.50 in / $12.50 out
Claude Sonnet 4.6	$3.00	$0.30 (hit) / $3.75 (5m write)	50% off	Batch: $1.50 in / $7.50 out
Claude Sonnet 5 ⭐	$2.00 (intro) / $3.00 standard	$0.20 hit / $2.50 5m write (intro)	50% off	Intro pricing through 2026-08-31 00:00 UTC: $2.00 / $10.00; standard from 2026-09-01 00:00 UTC: $3.00 / $15.00
Claude Sonnet 4.5	$3.00	$0.30 (hit) / $3.75 (5m write)	50% off	Batch: $1.50 in / $7.50 out
GPT-5.4	$2.50	$0.25	50% off	Data residency +10%
GPT-5.4 mini	$0.75	$0.075	50% off	—
GPT-5.4 nano	$0.20	$0.02	50% off	—
GPT-5.3-Codex	$1.75	$0.175	50% off	—
GPT-5.5 Instant	$0.20	$0.02	50% off	Free-tier ChatGPT default since 2026-05-05; API pricing $0.20/$1.00
Gemini 3.1 Pro	$2.00	$0.20–$0.40 + $4.50/hr storage	50% off	Tiered by input length
Gemini 3.5 Flash	$1.50	$0.15	50% off	Released 2026-05-19; 4x faster than 3.1 Pro
Gemini 3 Flash	$0.50	$0.05 + $1.00/hr storage	50% off	—
Gemini 3.1 Flash-Lite	$0.25	$0.025 + $0.25/hr storage	50% off	Cached input $0.025/1M (Vertex standard tier, ≤200K ctx)
DeepSeek-V4-Flash	$0.14	$0.0028 (hit)	—	Cache-hit pricing from 2026-04-26 12:15 UTC
DeepSeek-V4-Pro	$0.435	$0.003625 (hit)	—	75% discount made permanent (2026-05-31); list price now $0.435 / $0.87
DeepSeek-V3.2	$0.28	$0.028	—	No formal batch API
Qwen3.5-Max	$0.40	Available	50% off	—
Qwen3.7-Max	$2.50	$0.25	—	90% cached discount; proprietary API-only
GLM-5 / GLM-5.1	$1.00	$0.20	—	—
Grok 4	$3.00	$0.75	—	—
Grok 4 Fast	$0.20	$0.05	—	Aliased to Grok 4.3; output $0.50/1M
Grok 4.20	$1.25	$0.20	—	All variants: reasoning, non-reasoning, multi-agent
Grok 4.3	$1.25	$0.20	—	~38% cheaper input than 4.20; released 2026-05-01
Grok-4.5 ⭐	$2.00	$0.50 (hit)	—	Multimodal text+image; function calling, structured outputs, reasoning; released 2026-07-08
Mistral Small 4	$0.15	$0.015 (estimated)	—	119B MoE; multimodal; released 2026-03-16
Command A+ ⭐	$0.40	—	—	25B active (218B total MoE); faster than GPT-5.4 nano, Claude Haiku, Grok 4.3; released 2026-05
Mistral Large 3	$0.50	$0.05 (cached)	—	Cached input $0.05/1M; 256K context
Step-3.5-Flash	$0.10	—	—	—
Claude Fable 5	$10.00	$1.00 (hit) / $12.50 (5m write)	50% off	Batch: $5.00 in / $25.00 out; fast mode $20/$100
Claude Mythos 5	$10.00	$1.00 (hit) / $12.50 (5m write)	50% off	Same pricing as Fable 5; restricted access via Project Glasswing
MiniMax M2.7	$0.30	—	—	Self-evolution capabilities; Agent Teams native; SWE-Pro 56.22%; 205K context; released 2026-03-18
MiniMax M3 (≤512K)	$0.30	$0.06 (hit)	—	Extended context (>512K): $0.60/$2.40; released 2026-06-01
Kimi K3 ⭐	$3.00	$0.30 (hit, 90% cached)	—	2.8T MoE; 1M context; open weights promised by 2026-07-27; released 2026-07-16
Meta Muse Spark 1.1 ⭐	API-only	—	—	Closed-weights; 1M context; agentic tool and computer use; released 2026-07-09
Inkling ⭐	$1.87	—	—	975B/41B active MoE; 1M context; Apache 2.0; multimodal; released 2026-07-15
Qwen 3.8 ⭐	API-only	—	—	2.4T parameters; open-weight promised; preview on Token Plan/Qoder/QoderWork; released 2026-07-19
Qwen3.7-Plus	$0.40	—	—	Released 2026-06-03; 1M context
Step 3.7 Flash	Free	—	—	Free tier available; 1M context; released 2026-05-28
NVIDIA Nemotron 3 Ultra	Free	—	—	Free via NVIDIA NIM; 1M context (NVFP4 Blackwell)
GPT-4.1 ⭐	$2.00	$0.20	—	1M context; knowledge cutoff June 2024; retired from ChatGPT Feb 13, 2026; API still available
GPT-4.1 mini ⭐	$0.40	$0.04	—	1M context; retired from ChatGPT Feb 13, 2026; API still available
GPT-4.1 nano ⭐	$0.10	$0.01	—	1M context; retired from ChatGPT Feb 13, 2026; API still available; OpenAI's cheapest model
GPT-5.6 Sol ⭐	$5.00	$0.50 (90% cached) / $6.25 (cache write)	50% off	GA; cache write 1.25x input; cached read 90% discount; released 2026-07-09
GPT-5.6 Terra ⭐	$2.50	$0.25 (90% cached) / $3.125 (cache write)	50% off	GA; cache write 1.25x input; cached read 90% discount; released 2026-07-09
GPT-5.6 Luna ⭐	$1.00	$0.10 (90% cached) / $1.25 (cache write)	50% off	GA; cache write 1.25x input; cached read 90% discount; released 2026-07-09

Speed & Latency

Output throughput and time-to-first-token from Artificial Analysis and provider benchmarks.

Model	Output Speed (tok/s)	TTFT	Notes
Gemini 3.1 Flash-Lite	~250	~2.1s	Fastest budget Google model
Step-3.5-Flash	85–350	—	Variable by provider; peak ~350 tok/s
Gemini 3 Flash	~193	~4.16s	—
MiniMax-M2.5 Lightning	~100	—	Faster tier
GPT-5.3-Codex	~86	~77.86s	High TTFT due to extended reasoning
Grok 4	~56	~8.96s	—
MiniMax-M2.5 Standard	~50	—	—

Most frontier models (Claude Opus/Sonnet 4.6, GPT-5.4, Gemini 3.1 Pro, etc.) have not yet been benchmarked on Artificial Analysis as of April 2026. Anthropic notes that fast mode on Claude Opus 4.7 is deprecated and scheduled for removal on 2026-07-24 00:00 UTC; use Claude Opus 4.8 for current Opus fast-mode workflows. Source: 🔗

Training Data Cutoffs

Knowledge cutoff dates — the point after which a model has no training data.

Model	Training Cutoff	Notes
Claude Sonnet 4.6	Jan 2026	Most recent cutoff among frontier models
Claude Sonnet 5 ⭐	Jan 2026	Reliable knowledge cutoff Jan 2026; released 2026-06-30 00:00 UTC
Claude Opus 4.6	Aug 2025	Reliable knowledge: May 2025
GPT-5.4 / mini / nano	Aug 31, 2025	—
GPT-5.3-Codex	Aug 31, 2025	—
Grok 4 Fast	Jul 2025	—
Grok 4.20 / 4.3	Jul 2025	Approximate; not publicly disclosed
Grok-4.5 ⭐	~Nov 2024	Same cutoff family as Grok 4; approximate; not separately disclosed
DeepSeek-V4 (Flash/Pro)	May 2025	—
Gemini 3.1 Flash-Lite	Jan 2025	—
Gemini 3.1 Pro / 3 Pro / 3 Flash	Jan 2025	—
Grok 4	~Nov–Dec 2024	Approximate
DeepSeek-V3.2	Jul 2024	—
Llama 4 Scout / Maverick	Aug 2024	—
DeepSeek-R1	~Oct 2023	Based on base model
GPT-4.1 / mini / nano	June 2024	Retired from ChatGPT Feb 13, 2026; API still available
GPT-4.5	June 2024	Retired from API July 2025; retiring from ChatGPT 2026-06-27
Kimi K3 ⭐	~Jul 2025	Approximate; not separately disclosed; released 2026-07-16
Meta Muse Spark 1.1 ⭐	~Jan 2026	Approximate; not separately disclosed; released 2026-07-09
Inkling ⭐	~Jan 2026	Approximate; not separately disclosed; released 2026-07-15
Qwen 3.8 ⭐	~Jan 2026	Approximate; not separately disclosed; released 2026-07-19

Models not listed (Qwen, GLM, MiniMax, Kimi, Step, Mistral): training cutoff not publicly disclosed.

Multilingual Support

Model	Languages	Details
Qwen3.5-Max	201	Largest language coverage
Llama 4 Scout	200	Pre-training languages
Qwen3-Max-Thinking	119	Qwen3 series
Gemini 3 Flash	100	91.8% MMMLU score across 100 languages
Gemini 3.1 Pro / 3 Pro	100+	—
Gemini 3.1 Flash-Lite	100	91.3% MMMLU score
Llama 4 Maverick	12	Output languages
Claude (all)	Many	English-optimized; broad multilingual; Sonnet 5 supports multilingual capabilities and vision
GPT-5.4 (all)	Many	Broad multilingual coverage
DeepSeek (all)	Many	Chinese + English focused
Grok (all)	Many	—
GLM-5 / GLM-5.1	Many	28.5T token training data
Inkling ⭐	Many	Multimodal (text/image/audio); released 2026-07-15
Qwen 3.8 ⭐	Many	Multilingual support; released 2026-07-19

Structured Output & Function Calling

All frontier models support structured JSON output and function/tool calling except where noted.

Capability	Supported Models	Not Supported
Structured Output (JSON mode)	All models listed in Frontier table	Gemini 3 Deep Think (no API)
Function Calling / Tool Use	All models listed in Frontier table	Gemini 3 Deep Think (no API)

Gemini 3 Deep Think is available only via Gemini's in-app Think mode — no API access for structured output or function calling.

Regional Availability

Provider	API Availability	Cloud Partners	Notes
Anthropic	Global	AWS Bedrock, GCP Vertex AI	US-only inference at 1.1x via `inference_geo`
OpenAI	Global	Azure OpenAI	Data residency endpoints +10% (post-3/5/26)
Google	Global	Google AI Studio, Vertex AI	Some regional restrictions per Google terms
DeepSeek	Global	Azure (R1 only, select regions)	China-based servers
Alibaba (Qwen)	Global	Alibaba Cloud Model Studio	China-based; globally accessible
Zhipu AI (GLM)	Global	Z.AI API	MIT license enables self-hosting anywhere
MiniMax	Global	MiniMax API	—
Moonshot AI (Kimi)	Global	platform.kimi.ai	MIT open-weight
xAI (Grok 4 / 4.20 / 4.3 / 4.5 / 4 Fast)	US-focused	Oracle OCI (East/Midwest/West), Amazon Bedrock (June 2026)	Major price cuts: Grok 4.20 now $1.25/$2.50; Grok 4 Fast aliased to 4.3; Grok 4.3 on Bedrock at $1.25/$2.50; Grok-4.5 flagship at $2.00/$6.00 (500K context, text+image)
Mistral	Global	Azure AI Foundry, AWS, GCP	—
Meta (Llama)	Global (self-host)	All major cloud providers	Llama 4 Community License
Meta (Muse Spark 1.1)	US-only (public preview)	Meta Model API	Closed-weights frontier multimodal reasoning model; public preview for US developers; released 2026-07-09
Thinking Machines Lab (Inkling)	US-focused	Inkling API	Open-weight model (Apache 2.0); 975B/41B active MoE; 1M context; released 2026-07-15
Alibaba (Qwen 3.8)	Global (preview)	Token Plan / Qoder / QoderWork	Open-weight model; 2.4T parameters; preview access; released 2026-07-19
StepFun	Global	HuggingFace	Apache 2.0 open-source

Free-Source Models 🆓

Self-hostable models with permissive licenses or open weights for privacy, cost control, and customization.

Model	Company	Params	Context	License
DeepSeek-V4-Flash	DeepSeek	1.6T / 49B active (MoE)	1M	Open Weight
Qwen3.5-Max	Alibaba	397B / 17B active (MoE)	262K	Apache 2.0
Qwen3-Max-Thinking	Alibaba	1T+	128K	Apache 2.0
Qwen3.6-27B	Alibaba	27B dense	262K	Apache 2.0
Qwen3.5-122B	Alibaba	397B / 17B active (MoE)	262K	Apache 2.0
Qwen3.5-27B	Alibaba	27B dense	262K	Apache 2.0
Mistral Small 4	Mistral AI	119B / 6B active (MoE, 128 experts)	256K	Apache 2.0
Mistral Large 3	Mistral AI	675B / 41B active (MoE)	256K	Apache 2.0
Command A+	Cohere	25B active (218B total MoE)	128K	Apache 2.0
Llama 4 Scout	Meta	109B	10M	Community
Llama 4 Maverick	Meta	400B	128K	Community
GPT-OSS-120B	OpenAI	117B	128K	Apache 2.0
GPT-OSS-20B	OpenAI	21B	128K	Apache 2.0
Qwen3-Coder	Alibaba	480B	262K	Apache 2.0
GLM-5.1	Zhipu AI	754B / 40B active (MoE)	200K	MIT
GLM-4.7	Zhipu AI	400B+ MoE	205K	Open Weight
Gemma 4 31B	Google	31B dense	256K	Apache 2.0
Gemma 4 26B-A4B	Google	26B MoE (4B active)	256K	Apache 2.0
Gemma 4 E4B	Google	4B dense	256K	Apache 2.0
Gemma 4 E2B	Google	2B dense	256K	Apache 2.0
Gemma 4 12B	Google	12B dense	256K	Apache 2.0
Qwen3-Coder 7B	Alibaba	7B dense	128K	Apache 2.0
Qwen 2.5 Coder 32B	Alibaba	32B dense	128K	Apache 2.0
DeepSeek Coder-V2	DeepSeek	236B / 2.4B active	128K	MIT
Step-3.5-flash	StepFun	196B / 11B active (MoE)	256K	Open Weight
Yi-Coder	01.AI	9B/1.5B	128K	Apache 2.0
Lizzy-7B	Flower Labs	7B	—	MIT
MiMo-V2.5	Xiaomi	309B / 15B active	1M	MIT
MiMo-V2.5-Pro	Xiaomi	1.02T / 42B active	1M	MIT
NVIDIA Nemotron 3 Ultra	NVIDIA	550B / 55B active (MoE)	262K (BF16) / 1M (NVFP4 Blackwell)	Open Weight (NVIDIA)
MiniMax M3 ⭐	MiniMax	~428B / ~23B active (MoE)	1M	MiniMax Community
MiniMax M2 ⭐	MiniMax	230B / 10B active (MoE)	1M	Modified MIT
MiniMax M2.1 ⭐	MiniMax	229B / 10B active (MoE)	1M	Apache 2.0
GLM-5.2 ⭐	Zhipu AI	753B / 40B active (MoE)	1M	MIT
Kimi K2.7 Code ⭐	Moonshot AI	1T / 32B active (MoE)	262K	Modified MIT
MiniMax M2.7 ⭐	MiniMax	~400B+ (MoE)	205K	MiniMax Community
Kimi K2.6	Moonshot AI	1T / 32B active (MoE)	262K	Modified MIT
Kimi K3 ⭐	Moonshot AI	2.8T / 16 active (MoE)	1M	Modified MIT
LongCat-2.0 ⭐	Meituan	1.6T / ~48B active (MoE)	1M	MIT
Hy3 ⭐	Tencent	295B / 21B active (MoE)	256K	Apache 2.0
Ornith-1.0 ⭐	DeepReinforce	9B / 31B / 35B-MoE / 397B-MoE	—	MIT

Deployment Options

Local Inference Tools:

Ollama - Easy local deployment
LM Studio - User-friendly GUI
llama.cpp - Efficient CPU inference
vLLM - High-throughput serving
SGLang - Structured generation

Cloud Deployment:

Hugging Face Inference - Managed deployment
AWS SageMaker - Full control
Google Cloud Vertex - Integrated
RunPod - GPU rental

Coding Models 💻

Specialized AI models optimized for software development tasks.

SWE-bench Verified Leaderboard

Rank	Model	Company	SWE-bench Verified
🥇 #1	Claude Mythos 5 ⚠️	Anthropic	95.5%
🥈 #2	Claude Fable 5 ⚠️	Anthropic	95.0%
🥉 #3	Claude Mythos Preview	Anthropic	93.9%
#4	Claude Sonnet 5 ⭐	Anthropic	92.4%
#5	GPT-5.5 Pro	OpenAI	92.3%
#6	GPT-5.5	OpenAI	88.7%
#7	Claude Opus 4.8	Anthropic	88.6%
#8	Claude Opus 4.7	Anthropic	87.6%
#9	GPT-5.3-Codex	OpenAI	85.0%
#10	Ornith-1.0-397B ⭐	DeepReinforce	82.4%
#11	Claude Opus 4.5	Anthropic	80.9%
#12	Claude Opus 4.6	Anthropic	80.8%
#13	DeepSeek-V4-Pro (Max)	DeepSeek	80.6%
#13	Gemini 3.1 Pro	Google	80.6%
#14	MiniMax M3 ⭐	MiniMax	80.5%
#14	Qwen3.7-Max	Alibaba	80.4%
#15	Kimi K2.6	Moonshot AI	80.2%
#16	MiniMax-M2.5	MiniMax	80.2%
#17	GPT-5.2	OpenAI	80.0%
#18	Claude Sonnet 4.6	Anthropic	79.6%
#19	DeepSeek-V4-Flash (Max)	DeepSeek	79.0%
#20	Qwen3.6 Plus	Alibaba	78.8%
#21	Kimi K2.7 Code ⭐	Moonshot AI	78.2%
#22	Gemini 3 Flash	Google	78.0%
#22	MiMo-V2-Pro	Xiaomi	78.0%
#24	GLM-5	Zhipu AI	77.8%
#25	Mistral Medium 3.5	Mistral AI	77.6%
#26	Claude Sonnet 4.5	Anthropic	77.2%

Commercial Coding Models

Model	Developer	Pricing	Best For
Claude Fable 5 ⚠️ ⭐	Anthropic	$10.00 / $50.00 per 1M	Mythos-class coding; SWE-bench 95.0%; restored globally 2026-07-01 00:00 UTC with stronger safeguards
Claude Opus 4.8	Anthropic	$5.00 / $25.00 per 1M	Agentic coding, complex tasks, best available Opus
Claude Sonnet 5 ⭐	Anthropic	$2.00 / $10.00 per 1M intro through 2026-08-31 00:00 UTC	Agentic coding at Sonnet latency and price; 92.4% SWE-bench Verified; 63.2% SWE-bench Pro
Kimi K2.7 Code ⭐	Moonshot AI	$0.95 / $4.00 per 1M	Open-weight coding agent; 30% fewer thinking tokens vs K2.6; now available in GitHub Copilot model picker (July 1, 2026)
Claude Opus 4.6	Anthropic	$5.00 / $25.00 per 1M	Agentic coding, complex tasks
GPT-5.5 Pro	OpenAI	$30.00 / $180.00 per 1M	Highest benchmark coding
GPT-5.3-Codex	OpenAI	$1.75 / $14.00 per 1M	Agentic coding, 7+ hour autonomy
Claude Haiku 4.5	Anthropic	$1.00 / $5.00 per 1M	Low-latency coding, sub-agents, computer use
GLM-5-Code	Zhipu AI	$1.20 / $5.00 per 1M	Code generation, refactoring
MiniMax M3	MiniMax	$0.30 / $1.20 per 1M	Frontier coding + 1M context + native multimodal; open weights pending
MiniMax M2 ⭐	MiniMax	$0.30 / $1.20 per 1M	Open-weight MoE (230B/10B active); 1M context; 10x cheaper than Sonnet; released 2026-07-02
MiniMax M2.1 ⭐	MiniMax	$0.30 / $1.20 per 1M	Open-weight MoE (229B/10B active); 1M context; Apache 2.0; multilingual coding; released 2026-07-02
MiniMax M2.7 ⭐	MiniMax	$0.30 / $1.20 per 1M	Self-evolution; Agent Teams native; SWE-Pro 56.22%; 205K context; released 2026-03-18
MiniMax-M2.5	MiniMax	$0.30 / $1.20 per 1M	Code generation, refactoring
Claude Sonnet 4.5	Anthropic	$3.00 / $15.00 per 1M	Code review, refactoring
Mistral Small 4	Mistral AI	$0.15 / $0.60 per 1M	Unified reasoning + coding + multimodal; open-weight
Codestral	Mistral AI	$0.30 / $0.90	Real-time completion
North Mini Code	Cohere	Free (OpenRouter) / API	30B MoE (3B active); Apache 2.0; released 2026-06-09
Qwen3.7-Max	Alibaba	$2.50 / $7.50 per 1M	Agentic coding, long-horizon autonomy, reasoning
Grok 4.3	xAI	$1.25 / $2.50 per 1M	Flagship reasoning, lowest hallucination
MAI-Code-1-Flash	Microsoft	$0.75 / $4.50 per 1M	GitHub Copilot integration, fast everyday coding

Open-Source Coding Models

Model	Developer	License	Hardware
Kimi K2.7 Code ⭐	Moonshot AI	Modified MIT	80-160 GB VRAM
Kimi K3 ⭐	Moonshot AI	Modified MIT	160-320 GB VRAM
LongCat-2.0 ⭐	Meituan	MIT	320+ GB VRAM
GLM-5.2 ⭐	Zhipu AI	MIT	80-160 GB VRAM
MiniMax M2 ⭐	MiniMax	Apache 2.0	48-80 GB VRAM
MiniMax M2.1 ⭐	MiniMax	Apache 2.0	48-80 GB VRAM
MiniMax M3 ⭐	MiniMax	MIT	80-160 GB VRAM
MiniMax M2.7 ⭐	MiniMax	MiniMax Community	80-160 GB VRAM
Ornith-1.0 ⭐	DeepReinforce	MIT	160-320 GB VRAM
GPT-OSS-120B	OpenAI	Apache 2.0	80-160 GB VRAM
Qwen3-Coder	Alibaba	Apache 2.0	160-320 GB VRAM
DeepSeek-Coder-V2	DeepSeek	MIT	48-80 GB VRAM
GLM-5.1	Zhipu AI	MIT	80-160 GB VRAM
Phi-4	Microsoft	MIT	24-48 GB VRAM
Qwen3-Coder 7B	Alibaba	Apache 2.0	24-48 GB VRAM

Reasoning Models 🧠

Models optimized for step-by-step reasoning, mathematical problem-solving, and complex logical inference.

AIME 2025 Leaderboard

Rank	Model	AIME 2025	ARC-AGI-2	Notes
🥇 #1	MAI-Thinking-1 ⭐	97%	—	Microsoft's first in-house reasoning model; private preview
#2	Kimi K2.5 (Reasoning)	96.1%	—	Moonshot AI reasoning variant
#2	Kimi K2.5	96.1%	—	Open-weight reasoning
#4	GLM-4.7	95.7%	—	Zhipu AI open-weight
#5	MiMo-V2-Flash	94.1%	—	Xiaomi open-weight
#6	Claude Sonnet 4.5	87%	—	—
#7	Exaone 4.0 32B	85.3%	—	LG AI Research
#8	Nemotron 3 Nano Omni 30B	82.1%	—	NVIDIA
#9	LFM2.5-8B-A1B	42.5%	—	LiquidAI
#10	MiniCPM5-1B	40.4%	—	OpenBMB
#12	Step-3.5-Flash	97.3%	—	Best efficiency ratio
#12	GLM-5.2 ⭐	99.2% (AIME 2026)	—	Zhipu AI open-weight; leads AIME 2026 leaderboard
#13	Kimi K2.6	96.4%	—	Strong multimodal reasoning
#14	GLM-4.7	95.7%	—	Zhipu AI strong performer
#15	Claude Sonnet 4.6	~95%–95.6%	58.3%	Near-Opus performance
#16	GLM-5	92.7%	—	Thinking mode
#17	o4-mini	92.7%	—	Efficient OpenAI reasoning

Reasoning Model Details

Model	Type	Context	Pricing
Gemini 3 Deep Think	Reasoning	1M+	Ultra subscription
Qwen3-Max-Thinking	Reasoning/Coding	128K	$1.20 / $6.00
o3 / o1-Pro	Reasoning	128K	$2-150 / $8-600
GPT-5.5 Pro	Reasoning	1.05M	$30.00 / $180.00
Gemini 3 Pro	General/Multimodal	1M+	$2.00 / $12.00
DeepSeek-R1	Reasoning	128K	$0.50 / $2.15
Claude Sonnet 4.5	Hybrid	200K	$3.00 / $15.00
Claude Sonnet 5 ⭐	Hybrid	1M	$2.00 / $10.00 intro; $3.00 / $15.00 standard
GPT-Rosalind	Life Sciences Reasoning	128K	Pay-per-token (Research Preview)
MAI-Thinking-1	Reasoning	256K	TBD (Private Preview)

Use Cases

Mathematical Problem Solving: Qwen3-Max-Thinking, GPT-5.5 Pro, Gemini 3 Pro
Scientific Analysis: Claude Sonnet 5, Claude Opus 4.6, GPT-5.5, Gemini 3 Pro
Strategic Planning: o3/o1-Pro, Claude Sonnet 4.5, DeepSeek-R1
Code Debugging: Claude Sonnet 5, Claude Sonnet 4.5, GPT-5.3-Codex, DeepSeek-V3.2

Multimodal Models 🎨

Models capable of processing and generating multiple types of content: text, images, audio, and video.

Leading Multimodal Models

Model	Developer	Context	Key Features
GPT-5.4	OpenAI	1M	Unified multimodal, audio
Gemini 3 Pro	Google	1M+	Native multimodal, video
Claude Sonnet 4.5	Anthropic	200K	Document understanding
Claude Sonnet 5 ⭐	Anthropic	1M	Text and image input, 128K output, agentic coding and tool use
Llama 4 Maverick	Meta	128K	Open multimodal
MiniMax M3	MiniMax	1M	First open-weight model: frontier coding + 1M context + native image & video understanding
Kimi K3 ⭐	Moonshot AI	1M	Open-weight frontier model: 2.8T MoE, native vision, $3.00/$15.00 API pricing; released 2026-07-16
Meta Muse Spark 1.1 ⭐	Meta	1M	Closed-weights frontier multimodal reasoning model; agentic tool and computer use; API-only; released 2026-07-09
Gemini Omni Flash	Google	—	Native multimodal (text+image+audio+video input, video generation), any-to-any creation
Nemotron 3 Nano Omni	NVIDIA	30B (3B active)	Vision, audio, language unified, 9x throughput

Vision Capabilities

Model	MMMU / MMMU-Pro	MathVista	DocVQA
Gemini 3.1 Pro	95% (MMMU-Pro)	—	—
GPT-5.4	94% (MMMU-Pro)	—	—
Gemini 3 Pro	81% (MMMU-Pro)	—	—
Gemini 3 Flash	80% (MMMU-Pro)	—	—
Claude Sonnet 4.5	77.8% (MMMU)	—	—
Llama 4 Maverick	73.4% (MMMU)	—	—

Audio & Video

Model	Speech-to-Text	Text-to-Speech	Video Input
Gemini 3 Pro	✅	✅	✅
GPT-5	✅	✅	⚠️
Whisper v3	✅	❌	✅

Image Generation

Model	Developer	License	Best For
Gemini 3.1 Flash-Lite Image ⭐	Google	Proprietary	Gemini image model card updated 2026-06-30 00:00 UTC; image + text + audio understanding and generation; output images priced by resolution
Nano Banana 2 Lite ⭐	Google	Proprietary	Fastest Gemini Image model; 4s generation; GA 2026-06-30; $0.034 per 1K images; rolling out to Gemini app, Search, NotebookLM, Photos, Stitch, Flow, Ads
GPT Image 2	OpenAI	Proprietary	ChatGPT Images 2.0; improved text rendering, multilingual support, advanced visual reasoning; released 2026-04-21; API pricing varies by quality/resolution
MAI-Image-2-Efficient	Microsoft	Proprietary	Production-ready quality, 41% lower cost
Flux 2	Black Forest Labs	Apache 2.0 (Dev); Proprietary (Pro)	Released Nov 2025; exceptional photorealism + natural language; Flux 2 Pro/Dev/Schnell variants
Flux.1	Black Forest Labs	Apache 2.0	High-fidelity art (original Flux family)
Stable Diffusion 3.5	Stability AI	Community License	Fine-tuning; MMDiT architecture, 2.5B params
GLM-Image	Zhipu AI (Z.ai)	API	Fast image generation
CogView-4	Zhipu AI (Z.ai)	API	Creative image generation
Firefly AI Assistant	Adobe	Public Beta (2026-04-27)	Creative agent, 60+ tools, Photoshop/Premiere integration

Hardware Requirements 🖥️

Comprehensive hardware specifications for self-hosting AI models.

Quick Reference by Model Size

Model	Params	Q4 Size	Min VRAM	Rec VRAM	Min RAM
Phi-4	14B	8 GB	24 GB	48 GB	32 GB
GPT-OSS-20B	21B	12 GB	24 GB	48 GB	32 GB
Llama 4 Scout	109B	66 GB	48 GB	80 GB	96 GB
GPT-OSS-120B	117B	70 GB	80 GB	160 GB	128 GB
DeepSeek-Coder-V2	236B	143 GB	48 GB	80 GB	192 GB
Llama 4 Maverick	400B	242 GB	160 GB	320 GB	320 GB
DeepSeek-V4 family	671B	404 GB	80 GB	320 GB	512 GB
Qwen3-Max-Thinking	1T+	600+ GB	160 GB	640 GB	768 GB

By Hardware Tier

Consumer/Entry Level (24-48 GB VRAM):

Phi-4, GPT-OSS-20B, Yi-Coder, Qwen2.5-Coder
Recommended GPUs: RTX 3090 (24GB), RTX 4090 (24GB)

Professional (80-160 GB VRAM):

Llama 4 Scout, GPT-OSS-120B, DeepSeek-Coder-V2
Recommended GPUs: A100 80GB, 2x A100 40GB

Enterprise (320+ GB VRAM):

Llama 4 Maverick, GLM-4.7, DeepSeek-V4 family, Qwen3-Max-Thinking
Recommended GPUs: 4x A100 80GB, 8x A100 80GB

Quantization Explained

Level	Bits	Size vs FP16	Quality	Use Case
FP16/BF16	16	100%	Best	Training
Q8_0	8	~50%	Excellent	High-quality inference
Q4_K_M	4	~25%	Good	Recommended for deployment
Q3_K_M	3	~19%	Fair	Limited resources

Comprehensive Benchmark Reference 📈

Detailed benchmark scores across all major evaluations. Scores are percentages (%) unless noted. Arena Elo scores are integers. — = not publicly reported. Data as of 2026-07-19 16:54 UTC.

Full Benchmark Table

Model	GPQA Diamond	MMLU-Pro	Arena Elo (Text)	HLE	SWE-bench Verified	SWE-bench Pro	LiveCodeBench	AIME 2025	ARC-AGI-2	MMMU-Pro	IFEval	FrontierMath
Claude Opus 4.6	91.3%	—	1500	40.0–53.0%	80.8%	—	—	99.8%	68.8%	—	—	—
Claude Opus 4.8	93.6%	—	—	—	88.6%	69.2%	—	—	—	—	—	—
Claude Mythos Preview	94.5%	—	—	—	93.9%	—	—	—	—	—	—	—
GPT-5.5	93.6%	—	1495	42.1–55.0%	88.5%	—	—	99.9%	71.2%	—	—	52%
GPT-5.5 Instant	91.0%	—	—	—	88.7%	—	—	—	—	—	—	—
NVIDIA Nemotron 3 Ultra	—	—	—	—	—	—	—	—	—	—	—	—
Ornith-1.0-397B ⭐	—	—	—	—	82.4%	—	—	—	—	—	—	—
GPT-5.5 Pro	95.1%	96%	1520	48.5–62.0%	92.3%	—	—	100%	78.5%	—	97%	58%
Claude Sonnet 4.6	89.9%	—	~1438	33.2–49.0%	79.6%	—	—	~95%	58.3%	—	—	—
Claude Sonnet 5 ⭐	—	—	—	43.2% no tools / 57.4% tools	92.4%	63.2%	—	—	—	—	—	—
Claude Sonnet 4.5	83.4%	88.0%	—	—	77.2%	—	—	87–100%	—	—	—	—
Claude Fable 5	94.5%	—	—	—	95.0%	80.3%	29.3% (FrontierCode)	—	—	—	—	—
Claude Mythos 5	~94.1%	—	—	—	95.5%	—	—	—	—	—	—	—
Mistral Small 4	~71.2%	—	—	—	—	—	—	—	—	—	—	—
Mistral Medium 3.5	—	—	—	—	77.6%	—	—	86.3%	—	—	—	—
GPT-5.4	92.0%	94%	1484	36.6–41.6%	~80%	57.7%	84–88%	88%	73.3%	94%	—	50% (Pro)
GPT-5.4 mini	87.5%	—	—	—	—	54.4%	—	—	—	—	—	—
GPT-5.3-Codex	91.5%	—	—	—	—	56.8%	85%	—	—	—	—	—
GPT-5.2	92.4%	—	1479	35.2%	80.0%	55.6%	—	100%	52.9%	—	95.6%	~40.3%
Gemini 3.1 Pro	94.3%	92%	1494	44.4–51.4%	80.6%	54.2–72%	71%	100%	77.1%	95%	95%	—
DeepSeek-V4-Pro (Max)	90.1%	—	—	—	80.6%	55.4%	—	—	—	—	—	—
Gemini 3.5 Flash	~90.4%	—	—	—	55.1%	—	—	72.1%	83.6%	84.2%	40.2%	—
Gemini 3 Pro	91.9–93.8%	83%	1486	37.5%	76.2%	43.3%	49%	98–100%	31.1–45.1%	81%	88%	38%
Gemini 3 Flash	90.4%	72%	1474	33.7%	78.0%	44%	—	—	—	80%	85%	—
Gemini 3 Deep Think	~97%	81%	—	48.4%	~58%	63%	58%	—	84.6%	—	—	—
DeepSeek-V3.2	87.1%	85.0%	—	25.1%	67.8%	—	—	89.3%	—	—	—	—
DeepSeek-V4-Flash (Max)	88.1%	—	—	—	79.0%	52.6%	—	—	—	—	—	—
DeepSeek-R1	71.5%	84.0%	—	8.5%	49.2%	—	63.5%	70.0%	—	—	—	—
Qwen3.5-Max	89.3%	—	—	—	76.4%	—	—	91.3%	—	79%	—	—
Qwen3.7-Max	92.4%	—	—	41.4%	80.4%	—	—	—	—	—	—	—
Qwen3.6 Plus	90.4%	—	—	—	78.8%	—	—	—	—	—	—	—
Qwen3-Max-Thinking	86.1%	—	—	26.2%	—	—	—	—	—	—	—	—
GLM-5	82.0%	—	~1451	10.4%	77.8%	—	—	92.7%	—	—	—	—
MiMo-V2-Pro	87.0%	—	—	—	78.0%	—	—	—	—	—	—	—
GLM-5.1	—	—	—	—	~80.4% (est.)	—	—	—	—	—	—	—
Kimi K2.6	90.5%	87.1%	—	31.5–50.2%	80.2%	—	85.0%	96.4%	—	78.5%	—	—
Kimi K3 ⭐	93.5%	—	—	—	88.3%	—	—	—	—	—	—	—
Meta Muse Spark 1.1 ⭐	—	—	—	—	61.5%	—	—	—	—	—	—	—
GLM-5.2 ⭐	~91.2%	—	—	—	—	62.1%	76.8% (MCP-Atlas)	99.2%	—	—	—	—
MiniMax M3	~92.9%	—	—	—	—	—	—	—	—	—	—	—
MiniMax-M2.5	85.2%	—	—	—	80.2%	55.4%	—	86.3%	—	—	—	—
Step-3.5-Flash	83.1%	—	—	—	74.4%	—	86.4%	97.3%	—	—	—	—
Grok 4	~91.5%	91.5%	~1493	50.7%	—	—	—	100%	—	—	—	—
Grok 4.20	—	—	—	—	—	—	—	—	—	—	—	—
Grok 4.3	—	—	—	—	—	—	—	—	—	—	—	—
Grok-4.5 ⭐	—	—	—	—	—	—	—	—	—	—	—	—
Llama 4 Maverick	69.8%	80.5%	—	—	—	—	43.4%	—	—	—	—	—
Llama 4 Scout	57.2%	74.3%	—	—	—	—	32.8%	—	—	—	—	—

FrontierMath Scores

FrontierMath is a benchmark of 350 original, exceptionally challenging mathematics problems created by expert mathematicians (Epoch AI). Problems span number theory, analysis, algebraic geometry, and category theory. Tier 4 problems can take research mathematicians multiple days.

Model	Tiers 1–3	Tier 4	Source
GPT-5.4 Pro	50%	~36–38%	Epoch AI
GPT-5.2 Pro	~40.3%	31%	Epoch AI
Gemini 3 Pro	38%	19%	Epoch AI
GPT-5.1 Thinking	~25%	—	llm-stats

Notes on Limited Availability Models

Claude Mythos Preview is not publicly available and is only accessible through Project Glasswing, an invitation-only partner program for cybersecurity applications. Benchmarks for this model are not publicly disclosed as of May 2026.

Benchmark Glossary

Benchmark	Description	Source
GPQA Diamond	Graduate-level science questions (PhD difficulty)	Google Research
MMLU-Pro	Extended multi-task language understanding (harder than MMLU)	TIGER-Lab
Arena Elo	Crowdsourced human preference ranking	lmarena.ai
HLE	Humanity's Last Exam — expert-level questions	Scale AI
SWE-bench Verified	Real GitHub issue resolution (human-verified subset)	SWE-bench
SWE-bench Pro	More challenging subset of SWE-bench	SWE-bench
LiveCodeBench	Live competitive programming problems (not in training data)	LiveCodeBench
AIME 2025	American Invitational Mathematics Examination	MAA
ARC-AGI-2	Abstract reasoning challenge (fluid intelligence)	ARC Prize
MMMU / MMMU-Pro	Multi-discipline multimodal understanding	MMMU
IFEval	Instruction-following evaluation	Google Research
FrontierMath	Expert-level research mathematics (Epoch AI)	Epoch AI

Development Tools 🛠️

AI-powered tools for software development, from IDEs and CLI tools to API providers and IDE extensions.

IDEs 💻

Integrated Development Environments with built-in AI capabilities.

Agentic IDEs

IDE	Platform	Version	Release Date	Pricing	Key Features	GitHub
Firebase Studio	Web	-	-	Free (3 workspaces, up to 30 with Google Developer Program)	Cloud-based, Gemini, MCP	🔗
Lingma IDE (通义灵码)	Windows, macOS	-	-	Free (download)	Built-in agent, MCP tool use, terminal command execution	❌
Tonkotsu	Windows, macOS	-	-	Free (during early access)	Team of agents, workflow	🔗
OpenCode	Windows, macOS, Linux	-	-	Free (OSS)	Terminal, desktop, IDE extension, multi-provider	🔗
Codex app	Windows	-	2026-03-04 00:00 UTC	Included with Codex plans	Multiple agents, isolated worktrees, reviewable diffs, CLI and IDE interop	🔗
Visual Studio	Windows, macOS	17.14.12+, 18.1.0+	2026-01-06 00:00 UTC	Free / $250/yr	Gemini 3 Flash integration, faster performance, zero-migration upgrades, real-time profiler agent	❌
IntelliJ IDEA	Windows, macOS, Linux	2025.3.2	2026-01	Free / $149/yr	Java 24 support, Kotlin K2 mode, performance and memory improvements	❌
IBM Bob	Cross-platform	GA (April 28, 2026)	2026-04-28	Free trial + Enterprise plans	Multi-model orchestration, full SDLC, 45% productivity gain	🔗
PolyAI ADK	PolyAI	GA (April 22, 2026)	2026-04-22	Enterprise CX	AI-native dev, Cursor/Claude Code integration	🔗
Superset	Windows, macOS, Linux	-	2026-05-22	Free (OSS, YC P26)	Agentic IDE for running coding agents in parallel, git worktree management, remote workspaces, task tracking	🔗
Creor	Windows, macOS, Linux	-	2026	Free / Paid	AI-native IDE with specialized agents (Build, Plan, Explore), agent auto-routing, MCP support	🔗
Sinew	Windows, macOS, Linux	-	2026-05-13	Free (MIT)	Desktop AI coding harness, Tauri 2, multi-provider, agent swarm (2-8 agents), MCP, skills	🔗
JAT	Windows, macOS, Linux	-	2026-04-14	Free (MIT)	Self-contained agentic IDE, 20+ parallel agents, task management, unified environment	🔗
Qoder 1.0 ⭐	Windows, macOS, Linux	1.0	2026-06-16 00:00 UTC	Free / API	Alibaba Cloud Autonomous Development Desktop; Quest window for agent command; editor + agent parallel workspaces; GA 2026-06-16	🔗

Native AI Editors

Editor	Platform	Version	Release Date	Pricing	Key Features	GitHub
Zed	macOS, Windows, Linux	0.226.3	2026-03-03 00:00 UTC	Free (OSS) + Copilot $10/mo	Fast, collaboration, Gemini and Claude, Zeta AI, agent thread history, edit prediction providers, self-hosted OpenAI-compatible servers	🔗
Dyad	Windows, macOS, Linux	-	-	Free (OSS)	Local generation, BYO keys	🔗
Memex	macOS, Windows	-	-	Freemium (Free + $10/mo)	Agentic, browser↔desktop	🔗

VS Code Forks

IDE	Platform	Version	Release Date	Pricing	Autonomous	MCP	GitHub
Cursor ⭐	Windows, macOS, Linux	3.11	2026-07-10 00:00 UTC ⭐	Free / Pro $20/mo / Pro+ $60/mo / Ultra $200/mo; side chats, conversation search, and simpler project/repo pickers	✅	✅	❌
Windsurf	Windows, macOS, Linux	Quota plans	2026-03-19 00:00 UTC	Free / Pro $20/mo / Max $200/mo / Teams from $40 flex seat or $80 full user per month	✅	✅	❌
Trae	macOS, Windows	-	-	Free	❌	❌	🔗
PearAI	Windows, macOS, Linux	-	-	Free (OSS)	✅	❌	🔗
Void	Windows, macOS, Linux	-	-	Free (OSS)	✅	✅	🔗
Kiro ⭐	Windows, macOS, Linux	1.0.138	2026-07-13 00:00 UTC ⭐	Free preview; faster session startup, PowerShell trust on Windows, compaction-loop fix for long sessions, and improved MCP tool recovery; latest 1.0.x builds download directly while auto-updates roll out gradually	✅	✅	🔗
VS Code Agents	Windows, macOS, Linux	Insiders	2026-04-21	Free	✅	✅	🔗

Cursor vs. Windsurf Pricing (July 2026): Cursor Pro is $20/month, Pro+ is $60/month, and Ultra is $200/month. Windsurf Pro is $20/month and Max is $200/month, with Teams offering $40 flex seats or $80 full users per month. Cursor's current differentiator is broader cloud-agent and customization surface, while Windsurf's post-March 2026 quota plans simplify spend predictability. For a full alternatives analysis, see Cursor Alternatives & Better Value ↓.

Cursor Alternatives & Better Value 🎯

Cursor Pro is $20/month, while several official pricing pages and product docs show lower-cost or BYOK alternatives with similar agent workflows, local-model support, or broader IDE coverage.

Pricing Reality Check (July 2026)

Tool	Free Tier	Pro / Paid	Premium / Max	Notes
Cursor	2K completions + 50 premium req	$20/mo (Pro) / $60/mo (Pro+)	$200/mo (Ultra)	Heavy users report $40–60/mo effective cost; quotas tightened June 2025
Claude Code	Limited	$20/mo (Pro)	$100/mo (Max 5x) / $200/mo (Max 20x)	Significantly higher limits than Cursor Pro at same price
GitHub Copilot	Limited AI Credits	$10/mo base + $5 included AI Credits (Pro)	$39/mo base + $31 included AI Credits (Pro+) / $39/user (Enterprise)	Usage-based billing since June 1, 2026; bundled credits are consumed by chat, agent mode, cloud agent, CLI, and apps
Cline (BYOK)	Free (bring API key)	Free (pay API costs ~$5–15/mo)	Unlimited	5M+ installs; zero subscription cost
Windsurf	Light quota	$20/mo (Pro) / Teams from $40 flex seat or $80 full user	$200/mo (Max)	Quota-based plans replaced older credit plans in March 2026
Continue.dev	Free (OSS)	Free (BYOK)	Unlimited	No subscription; supports Ollama, LM Studio, all cloud APIs
Void	Free (OSS)	Free (BYOK)	Unlimited	VS Code fork + MCP; privacy-first, no telemetry
Aider	Free (BYOK)	Free (BYOK)	Unlimited	Git-native; used by serious OSS contributors
OpenHands	Free (OSS)	Free (self-host)	Unlimited	Full SDLC agent; MCP support

Value-For-Money Ranking

Rank	Tool	$/Month (Real)	SWE-bench	Limits	Verdict
🥇 1	GitHub Copilot Pro	$10 base + $5 credits	55%	Usage-based AI Credits (with bundled allowance)	Best value for light-moderate users
🥈 2	Cline (BYOK)	~$5–15 (API)	80.8%	Unlimited	Best value for power users
🥉 3	Windsurf Pro	$20	—	Quota-based plan	Best value paid Pro tier
4	Claude Code Pro	$20	80.8%	Higher than Cursor Pro	Best autonomy at $20/mo
5	Continue.dev	$0	—	Unlimited	Best free option
6	Cursor Pro	$20–60 (effective)	Unpublished	Caps hit easily	Good UX, poor value at scale
7	Claude Code Max	$100	80.8%	5× limits	Outclasses Cursor Ultra ($200)

Official Source Basis

This comparison uses official pricing pages, official changelogs, official extension marketplaces, and official GitHub repositories. Unofficial social feedback is not used for eligibility, pricing, or benchmark claims.

Why Specific Alternatives Beat Cursor

🥇 Cline (BYOK) — Best Overall Value Free VS Code extension with zero subscription cost. You pay only for your API usage (~$5–15/month for typical developer workloads). Achieved 80.8% SWE-bench Verified — identical to Claude Code. Supports 100+ providers including local Ollama/LM Studio models, meaning you can run it for $0/month with a local model like Qwen3-Coder or DeepSeek-Coder. Over 5 million installs make it the most-adopted free agent tool. Full MCP support.

🥈 Claude Code — Best Paid Alternative Anthropic's purpose-built autonomous coding CLI. At $20/mo (Pro), usage limits are substantially higher than Cursor Pro at the same price. Max 5× ($100/mo) offers more headroom than Cursor Ultra ($200/mo) at half the cost. Claude Code produces code described as ~30% less rework vs. Cursor. Won 67% of blind quality comparisons against Cursor. Integrates natively into VS Code and JetBrains via extension, plus terminal. Multi-session support, computer use (research preview), sub-agents.

🥉 GitHub Copilot — Best for Teams & IDE-Native Work At $10/month base with $5 in bundled AI Credits (Pro), GitHub Copilot requires no fork, no new editor, and no learning curve. Works inside existing VS Code, JetBrains, Vim, Neovim, and more. Since June 1, 2026, billing switched to usage-based AI Credits; chat, agent mode, cloud agent, CLI, and apps all consume the same credit pool. Pro+ starts at $39/month base with larger included credits. Best for developers who want AI assistance without changing their workflow.

Continue.dev — Best Free Fully-Featured Option 100% open-source, works in VS Code and JetBrains. BYOK means no subscription cost ever. Supports Ollama, LM Studio, and all major cloud APIs. Chat, autocomplete, and agent mode. Active development community. No telemetry by default.

Void — Best Privacy-First Fork Open-source VS Code fork with full MCP support, BYOK, and no telemetry. Feature-comparable to Cursor with zero subscription cost. Best for teams that need data sovereignty or work in air-gapped environments.

🎯 Bottom Line for Sharif's Use Case: If you use LM Studio and prefer local models, Cline or Continue.dev (both BYOK, both free) give you unlimited usage at zero subscription cost. For cloud-based work, Claude Code Pro ($20) remains the strongest autonomy-first alternative at the same entry price as Cursor Pro ($20). GitHub Copilot is still the easiest low-friction option if you want to stay inside your current IDE and manage against a usage-credit budget.

Web-Based IDEs

IDE	Platform	Version	Release Date	Pricing	Self-Hostable	Best For	GitHub
Replit 3	Web	-	-	Free Starter, Core $20/mo, Pro $100/mo	❌	Learning/Prototyping	❌
Bolt.new	Web	-	-	Free, Pro $20-25/mo, Teams $30/user/mo	❌	Quick apps	❌
Bolt.diy	Self-hosted	-	-	Free (MIT), bring your own API	✅	Self-hosted	🔗
Lovable	Web	-	-	Free (5 credits/day), Pro $25/mo, Business $50/mo	❌	UI/Full-stack	❌
v0	Web	-	-	Free ($5 credits/mo), Premium $20/mo, Teams $30/user	❌	React components	❌
Gitpod	Web	-	-	Free + Paid	❌	Cloud dev environments	❌
Rork	Web	-	-	Free & Paid (credits)	❌	Mobile apps (iOS/Android)	❌
Google Stitch	Web	-	2026-03	Free (Google account, 550 gen/mo)	❌	UI design, Figma/React export	❌
Google Antigravity 2.0	Windows, macOS, Linux (CLI), Web (IDE)	2.0	2026-05-19 00:00 UTC	Google AI Pro $19.99/mo / Ultra $99.99/mo (5×) or $200/mo (20×)	Agent-first platform: desktop app + CLI (`agy`) + SDK + Managed Agents API; multi-agent orchestration, parallel tasks, voice control	🔗
Jules	Web	-	2025-05-20 00:00 UTC	Free beta, higher limits on Google AI Pro / Ultra	Async repo agent, reviewable diffs, GitHub integration	❌

CLI Tools 🖥️

Command-line AI tools for autonomous coding and terminal enhancement.

Autonomous Coding Agents

Tool	Platform	Pricing	Key Features	GitHub
Aider	Windows, macOS, Linux	Free	Gold standard, Architect mode, thinking tokens	🔗
Claude Code 2.2.1+	macOS, Linux, Windows	Free + API	Sonnet 5 default support via Claude Code 2.1.197+; Opus 4.8 support; Fable 5 access restored globally 2026-07-01 00:00 UTC; simple mode file editing, multi-session support, `/autofix-pr`, plan mode, computer use (research preview), CLAUDE.md skills & sub-agents	🔗
Codex CLI ⭐	Windows, macOS, Linux	Included with ChatGPT Plus/Pro/Business/Edu/Enterprise or API key	Rust-native local coding agent; sandbox, approval modes, subagents, command autocomplete, workspace search; v0.144.5 (2026-07-16) improved dangerous-command detection and denial messaging; v0.144.4 (2026-07-14) had no user-facing changes; v0.144.2 restored Guardian auto-review policy	🔗
Junie CLI ⭐	Windows, macOS, Linux	Free (BYOK)	LLM-agnostic, JetBrains IDE integration (ACP), GA 2026-06-17, agentic debugging, PR review, local model support (LiteLLM/LMStudio/Ollama)	🔗
Goose	Windows, macOS, Linux	Free (Apache-2.0)	MCP, extensible, desktop app, 25+ providers	🔗
Grok Build	Windows, macOS, Linux	$300/mo (SuperGrok/X Premium Plus)	Plan mode, Arena mode, 2M context window, security-focused	❌
GPT-Pilot	Windows, macOS, Linux	Free	Full dev team simulation	🔗
OpenHands	Windows, macOS, Linux	Free	Cloud agents, MCP	🔗
Mentat	Windows, macOS, Linux	Free	Multi-file coordination	🔗
SERA	Linux, macOS	Free (Apache 2.0)	Open-source coding agent, 200K synthetic trajectories	🔗
Kimi K2.7 Code ⭐	Windows, macOS, Linux	$0.95 / $4.00 per 1M	Coding-focused agent; 30% fewer thinking tokens vs K2.6; Modified MIT; 262K context; released 2026-06-12	🔗
Kilo for GitHub ⭐	GitHub (Cloud)	Free (Kilo credits)	@kilocode-bot coding agent in GitHub issues/PRs; Cloud Agent backend; reads full thread context; released 2026-06-18	🔗
Hermes Agent v0.17.0 ⭐	Windows, macOS, Linux	Free (OSS)	Major desktop app update; live subagent watch-windows; VS Code theme support; 199K GitHub stars; released 2026-06-19	🔗
Kimi Code CLI	Windows, macOS, Linux	Free (MIT)	TypeScript CLI by Moonshot AI; successor to kimi-cli; powered by Kimi K2.6 (1T MoE); subagent support (coder/explore/plan); 262K context; SWE-bench Pro 58.6%; ACP (Zed, JetBrains); released 2026-06-06	🔗
MiMo Code ⭐	Windows, macOS, Linux	Free (MIT)	Xiaomi open-source terminal coding agent; 62% SWE-bench Pro; integrated with Claude Code, OpenClaw, Hermes; released 2026-06-11	🔗
Vibe (Mistral)	Windows, macOS, Linux	Free (Le Chat) / API	Agentic coding + work agent, multi-model Mistral harness, VS Code extension, Code Mode web surface, parallel sessions	🔗
Google Agents CLI ⭐	Windows, macOS, Linux	Free	Unified ADLC for Google Cloud Agent Platform; scaffold/run/evaluate/deploy agents; released 2026-04-22	🔗
Peri ⭐	Windows, macOS, Linux	Free (Apache-2.0)	Rust terminal coding agent; Claude Code compatible; multi-LLM; MCP & ACP support; 95%+ prompt cache; released 2026-03-20	🔗
AI Dev Kit	Cross-platform	Free	59 skills, 33 agents, TDD, security audit, CI/CD	🔗

Assisted CLI Tools

Tool	Developer	Pricing	Best For	GitHub
Gemini CLI ⚠️	Google	Free (now shut down)	Google ecosystem & Gemini models in-terminal; stopped serving requests June 18, 2026 for AI Pro/Ultra/free users; migrate to Antigravity CLI	🔗
Antigravity CLI (`agy`) ⭐	Google	Free (~20 req/day) / Pro/Ultra higher limits	Agentic coding, multi-agent orchestration, Go-based replacement for Gemini CLI; released 2026-06-18	🔗
Google Colab CLI ⭐	Google	Free (Colab account)	Run Python on remote Colab GPUs/TPUs from local terminal; AI agent compatible; released 2026-06-06	🔗

⚠️ Anthropic June 15 Billing Split: As of 2026-06-15, Anthropic moved Claude Agent SDK, claude -p, and Claude Code GitHub Actions to separate metered credits billed at full API rates. Pro subscribers get $20/mo credit; Max 20x gets $200/mo. Interactive Claude Code and Claude.ai remain on standard subscription. Fable 5 was included on Pro/Max/Team/Enterprise subscription plans at no extra cost through June 22; after that date, usage credits are required.

Gemini CLI (deprecation & security notes): As of June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions have stopped serving requests for Google AI Pro, Ultra, and free Gemini Code Assist for individuals, per Google's official announcement. Google ecosystem users should migrate agentic workflows to Antigravity CLI (agy). Enterprise Code Assist Standard/Enterprise licensees retain access. Current CLI generations add offline search, bundled ripgrep, optional color-accessible themes, interactive shell tool invocation, and an A2A-style agent registry surface in the tool (see Gemini CLI changelogs). Treat upgrades through v0.39.1+ as mandatory maintenance: public advisories described a critical remote-code path affecting earlier builds (widely reported as CVSS 10.0); confirm your installed build against the release notes for your package channel. Source: Google Developers Blog

CLI Tools by Programming Language

AI coding CLI tools categorized by their primary language support. All tools below accept plain English prompts.

Tool	Primary Languages	Multi-Language	Local LLM	Cloud API	Pricing	GitHub
Aider	Python, JS, TS, Go, Rust, Ruby, Java, C/C++	✅ (100+ langs)	✅ (Ollama, LM Studio)	✅	Free (OSS)	🔗
Claude Code	All (polyglot)	✅	❌	Claude API ($3–$15/M)	Free tool / API cost	❌
Codex CLI	Python, JS, TS, Bash	✅	❌	OpenAI API	Free OSS / API cost	🔗
OpenHands	Python, JS, TS, Go, Rust, Java	✅	✅	✅	Free (OSS)	🔗
Goose (Block)	Polyglot (25+ providers)	✅	✅ (Ollama, LM Studio)	✅	Free (OSS)	🔗
Continue	Polyglot (VS Code, JetBrains)	✅	✅ (Ollama, LM Studio)	✅	Free (OSS)	🔗
Qwen Code	Python, JS, TS, Go, Java	✅	✅ (Qwen models)	✅	Free (OSS)	🔗
Devstral CLI	Python, JS, TS, Go, Rust	✅	✅	Mistral API	Free OSS model / API cost	❌
OpenCode	Polyglot	✅	✅	✅	Free (OSS)	🔗
Mentat	Python, JS, TS, Go	✅	❌	OpenAI API	Free (OSS)	🔗
Amp (Sourcegraph)	All	✅	❌	✅	Free / Enterprise	❌

CLI for Programming Languages & Multiple Use

Purpose-built CLI tools for coding across specific languages or polyglot multi-stack workflows.

Tool	Language Focus	Platform	Pricing	Key Features	GitHub
Aider	Polyglot (Python, JS, TS, Go, Rust, any)	All	Free (BYOK)	Git-native multi-file edits, Architect mode, repo maps, thinking tokens	🔗
Claude Code	Polyglot	All	Free + API	Computer use, sub-agents, CLAUDE.md skills, Opus 4.7, multi-session	🔗
Codex CLI	Python, JS, TS	All	Free (OpenAI account)	Sandbox execution, approval modes, OpenAI models	🔗
OpenHands	Python, JS, TS, Go, Rust	All	Free (OSS)	Full SDLC agent, MCP, local LLM via Ollama	🔗
Goose	Polyglot	All	Free (Apache-2.0)	25+ providers, MCP, extensible extensions, desktop app	🔗
Continue	Polyglot	All	Free (OSS)	VS Code + JetBrains, custom models via Ollama/LM Studio	🔗
Qwen Code	Python, JS, TS, Go	All	Free	Optimized for Qwen3-Coder 480B, Apache 2.0	❌
Mentat	Polyglot	All	Free	Multi-file coordination, context-aware diffs	🔗
AI Dev Kit	Polyglot	All	Free	59 skills, 33 agents, TDD, security audit, CI/CD pipeline	🔗
Devstral CLI	Python, JS, TS, Go	All	Free (Mistral free tier)	Mistral's open coding model, OpenRouter free access	❌
Junie CLI	Polyglot	All	Free (BYOK)	LLM-agnostic, JetBrains IDE integration, MCP	🔗
SERA	Python, JS, TS	Linux, macOS	Free (Apache 2.0)	Open-source coding agent, 200K synthetic trajectories	🔗

Terminal Enhancers

Tool	Platform	Pricing	Key Features
Warp Terminal	macOS, Linux, Windows	Free	AI Agents, workflow sharing, Warp Drive automation
Microsoft Intelligent Terminal	Windows	Free (MIT)	AI agent pane (Agent Client Protocol), automatic error detection, GitHub Copilot CLI default, Windows Terminal fork — 2026-06-02 ⭐
Starship	Cross-platform	Free (OSS)	Fast cross-shell prompt; rich plugins & AI-adjacent ecosystem
Fig	macOS, Linux	Free	Autocomplete, AI suggestions
Google Colab CLI ⭐	Cross-platform	Free (Colab account)	Run Python on remote Colab GPUs/TPUs from local terminal; AI agent compatible; released 2026-06-06

IDE Add-ons 🧩

Extensions and plugins that add AI capabilities to existing IDEs.

Universal (Cross-Platform)

Add-on	Platform	Pricing	Context	Best For	GitHub
GitHub Copilot ⭐	VS Code, JetBrains, Vim	Usage-based AI Credits since June 1, 2026	Large	General coding; Kimi K2.7 Code available in model picker (July 1, 2026); Copilot Vision GA (July 1, 2026); enterprise session streaming/API, AI credit pools, and Copilot-only Gemini 2.5 Pro / Gemini 3 Flash deprecation on 2026-07-31	🔗
Supermaven	VS Code, JetBrains, Neovim	Free / $10/mo	1M	Large codebases	❌
Codeium	VS Code, JetBrains, Vim	Free / $15/mo / $60/mo	Medium	Free alternative	❌
Continue	VS Code, JetBrains	Free (OSS)	Custom	Self-hosted	🔗
Cody	VS Code, JetBrains, Web	Free (discontinued) / Enterprise Starter $19/mo / Enterprise $59/mo	Enterprise	Code search	🔗
Tabnine	VS Code, JetBrains, VS, Eclipse	Free / $39/mo	Local	Privacy	❌
Tabby	VS Code, JetBrains, Vim, Neovim	Free (OSS)	Self-hosted	Self-hosted code completion	🔗

VS Code Specific

Add-on	Pricing	Autonomous	MCP	Best For	GitHub
Codex	Free (with ChatGPT Plus $20/mo or Pro $200/mo)	✅	✅	OpenAI's official coding agent	🔗
Cline	Free	✅	✅	Full agent	🔗
GitHub Copilot (Agent Mode) ⭐	Usage-based AI Credits since June 1, 2026	⚠️	❌	Guided agent workflows; auto-opens PRs with self-review; Kimi K2.7 Code available; Copilot CLI in GitHub Actions now supports `GITHUB_TOKEN` with `copilot-requests: write`	🔗
RooCode	Free/Pro	⚠️	❌	Complex tasks	🔗
JetBrains AI Assistant	$10/mo (Pro)	✅	❌	JetBrains-quality AI now in VS Code (public preview); multi-file edits, Mellum LLM	❌
ReSharper for VS Code	Included with JetBrains subscription	❌	❌	C# code analysis + refactoring inside VS Code/Cursor; released 2026-03-05	❌
Keploy	OSS/Enterprise	❌	❌	Testing	❌
Microsoft Foundry Toolkit for VS Code ⭐	Free (Azure account)	✅	✅	End-to-end Hosted Agent lifecycle; GA 2026-04-16; consolidated AI Toolkit + Foundry extension; LangGraph samples; released 2026-06-02	🔗
VS Code ACP Client ⭐	Free (MIT)	✅	✅	Agent Client Protocol client; connects 11+ AI agents (Claude Code, Codex, Cursor, etc.) to VS Code; released 2026-06-10	🔗
Prompt Foundry ⭐	Free	❌	✅	VS Code/Cursor extension; modular prompt blocks with Liquid syntax; MCP server for context updates; released 2026-06-18	🔗
GrackerAI ⭐	Free	❌	❌	Validates llms.txt, robots.txt, AI visibility signals in VS Code; released 2026-06-16	❌
A11yResolver ⭐	Free (beta)	❌	❌	VS Code extension for accessibility remediation; AI agent flags WCAG issues; released 2026-06-18	❌

Claude Tag (Anthropic, June 2026): Not an IDE extension but worth noting here. Claude Tag brings an always-on AI teammate to Slack. Each channel gets an isolated Claude identity with persistent memory. Available on Team and Enterprise plans. Source: 🔗

JetBrains Specific

Add-on	Pricing	Claude Agent	Junie (GA)	Best For
JetBrains AI Assistant	$10/mo (Pro), $249/yr (Ultimate)	✅	✅	Deep IDE integration; Junie GA 2026-06-17
JetBrains Claude Agent	Included in subscription	✅	—	Native agent
JetBrains Junie ⭐	Included in subscription	—	✅	GA 2026-06-17; ACP-based IDE integration; agentic debugging; BYOK; local model support; IntelliJ/WebStorm/PyCharm/GoLand/CLion

JetBrains Marketplace Security (June 2026): 15 malicious AI plugins removed from JetBrains Marketplace; 7 publisher accounts terminated; remote kill-switch triggered. If you installed any affected plugins before June 17, 2026, revoke exposed API keys immediately.

API Providers 🔌

Services for accessing AI models via API. This pass prioritizes the cheapest verified public text APIs and the free or trial access that is explicitly documented by each provider.

Model Labs (Direct)

Provider	Cheapest verified public text model	Direct price	Free or trial access	Notes	Source
Groq	Llama 3.1 8B Instant	$0.05 input / $0.08 output	GroqCloud Free plan ($0)	Lowest verified public direct text API price in this pass	🔗, 🔗
Google Gemini	Gemini 2.5 Flash-Lite	Free tier: free of charge; paid tier $0.10 / $0.40	Gemini Developer API free tier	Batch and Flex drop to $0.05 / $0.20	🔗
DeepSeek	DeepSeek-V4-Flash	$0.14 input / $0.28 output	No standing free tier listed on pricing page	Cache-hit input is $0.0028	🔗
Mistral AI	Mistral Small 4	$0.15 input / $0.60 output	Free mode enabled by default, no credit card required	Good low-cost coding and multimodal option	🔗, 🔗
Cohere	Command R	$0.15 input / $0.60 output	Trial API keys are free but limited; Command A+ is free until rate limits are reached	Low-cost RAG and tool use option	🔗, 🔗, 🔗
OpenAI	GPT-5.4 nano	$0.20 input / $1.25 output	OpenAI quickstart includes one free test API request	Batch and Flex drop to $0.10 / $0.625	🔗, 🔗
Anthropic	Claude Haiku 4.5	$1.00 input / $5.00 output	New users receive a small amount of free credits	Lowest-cost Claude API model	🔗, 🔗
xAI	Grok-4.5	$2.00 input / $0.50 cached / $6.00 output	No public free tier listed on pricing page	New flagship (2026-07-08); 500K context, text+image; Grok 4.3 remains $1.25 / $2.50	🔗
Meta	Meta Muse Spark 1.1	API-only (public preview)	US-only public preview via Meta Model API	Closed-weights frontier multimodal reasoning model; agentic tool and computer use; 1M context; released 2026-07-09	🔗

Unified APIs & Aggregators

Provider	Models	Key Features
OpenRouter	400+	Unified API, crypto/fiat, model rankings, fallback routing, Speech & Transcription APIs, Model Fusion (launched 2026-06-12), private models, enterprise workspace controls; new models: Claude Opus 4.8, Step 3.7 Flash, MiniMax M3, Qwen3.7-Plus, NVIDIA Nemotron 3 Ultra, Microsoft MAI models, Nano Banana 2, Nano Banana Pro, GLM 5.2, Cohere North Mini Code (free), Nex-N2-Pro (free), GPT-Image-1.5, GPT-Image-2, Poolside Laguna, Qwen3-Coder (June 2026)
Hugging Face	Thousands	Serverless inference, free tier
LiteLLM	100+	Open-source proxy; single OpenAI-compatible API for any provider
Portkey	200+	Gateway: load balance, fallbacks, caching, observability

Model Router Services 🔀

Services that aggregate multiple models through a unified API, often with load balancing, caching, and cost optimization features.

Provider	Models	Pricing	Key Features
Together AI	200+	$0.10-$2.50/1M	Broadest open-weight catalog (text/image/video/audio/embeddings), NVIDIA Blackwell-optimized
Fireworks AI	50+	$0.10-$1.50/1M	Production-grade open-source inference, 90% cheaper than closed providers, NVIDIA Blackwell
Groq	20+	$0.05-$0.90/1M	Fastest LPU inference: 476 tok/s on 120B, <100ms TTFT, deterministic latency
Cerebras	10+	$0.10-$1.00/1M	Wafer-scale engine: ~3,000 tok/s, 5× faster than NVIDIA Blackwell, 80–150ms TTFT
NVIDIA NIM	91+	Free (cloud endpoints) / self-host	Broadest model variety: LLMs + vision + audio + bio + climate; self-hostable Docker containers
Vercel AI Gateway ⭐	100+	$5/mo free credits; paid credits at provider list rates, zero markup	Available on all plans; unified API for text, image, video, realtime, speech, embeddings, and reranking; BYOK on paid tier; model fallbacks, provider preferences, budgets, observability, and Routing Rules beta
Mercury 2 (Inception Labs) ⭐	1	$0.25/$0.75/1M	Diffusion LLM; 1,009 tok/s on Blackwell; tunable reasoning; 128K context; released 2026-05-12
Anyscale	100+	$0.20-$2.00/1M	OpenAI-compatible, HIPAA/SOC 2/EU data residency, enterprise contracts
Replicate	50+	$0.20-$2.00/1M	Model-as-a-service, versioning, GPU rental
OctoAI	20+	$0.25-$2.00/1M	High-performance inference
Predibase	12+	$0.30-$2.50/1M	Fine-tuning + multi-LoRA serving
RunPod Serverless	10+	$0.20-$2.00/1M	GPU serverless, flexible pricing
CoreWeave	8+	Custom	Enterprise-grade, dedicated GPUs
Infron ⭐	100+	Up to 35% off direct pricing	Enterprise-grade unified API; provisioned throughput plan; broad model selection; released 2026-06-10
Aethir Mesh ⭐	10+ open-source	Pay-per-use	Decentralized GPU infrastructure; single API for open-source models (DeepSeek V4, Kimi K2.6, GLM-5.1, etc.); released 2026-06-12

Budget Model Aggregators

Provider	Models	Pricing	Best For
DeepInfra	50+	Free tier + $0.10-1.00/1M	Free tier, diverse models
TensorZero	30+	Free tier + $0.15-1.50/1M	Cost optimization, caching
Scale AI	25+	Custom pricing	Enterprise, managed services
AssemblyAI	15+	Free tier + pay-per-use	Speech + text models

GPU Clouds

Provider	Type	A100 80GB ($/hr)	H100 80GB ($/hr)	Best For
Vast.ai	GPU Marketplace	~$0.09–$0.59	~$0.50–$1.50	Cheapest market pricing, diverse GPU options
RunPod	GPU Rental	$0.34–$0.69 (RTX 4090) / $0.60–$1.19 (A100)	$1.99–$3.29	Flexibility, per-second billing, cost-effective fine-tuning & inference
Lambda Labs	Cloud GPU	$1.29–$2.79	$2.49–$3.99	Reliable on-demand, academic discounts, 1-Click Clusters
DigitalOcean	GPU Droplets	$1.29	$2.99	Simple fine-tuning workflows
Vultr	Global Cloud	$1.29–$2.80	$1.99–$2.99	Hourly GPU instances, global regions
Hyperbolic	Decentralized	~$1.50	~$3.50	Crypto/Fiat payments
Cerebrium	Serverless GPU	~$2.10	~$3.40	Python-native ML inference & fine-tuning
Modal Labs	Serverless GPU	~$2.50–$3.40	~$3.95	Fine-tuning with LoRA, distributed training; per-second billing
Replicate	Model-as-a-Service	$5.04	$5.49	Quick deployment, serverless inference
Together AI	AI-Native Cloud	—	—	Fast inference & fine-tuning for open models
Fireworks AI	Inference & Fine-tuning	—	—	Fast inference, RFT for model shaping
CoreWeave	Enterprise GPU Cloud	~$2.70	~$6.16	Enterprise-grade, dedicated GPUs, Kubernetes-native
Databricks Mosaic AI	Integrated ML Platform	—	—	Enterprise fine-tuning, governed serving, RAG
NVIDIA DGX Cloud	Managed AI Training	Custom	Custom	Co-engineered clusters, maximum ROI for training

GPU pricing as of July 2026. RunPod/Vast.ai prices vary by community vs. secure cloud and spot vs. on-demand. RTX 4090 pricing from RunPod starts at $0.34/hr. Vast.ai A100 40GB from $0.52/hr. Modal Labs production workloads cost 3.75x base rates (3x non-preemptible multiplier x 1.25x US regional multiplier). CoreWeave A100/H100 pricing normalized from 8-GPU nodes. Lambda Labs A100 40GB at $1.29/hr, A100 80GB at $2.49-$2.79/hr, H100 SXM at $3.99/hr (8x) or $4.29/hr (1x). Vast.ai prices are market-set medians.

Inference Clouds

Provider	Specialization	Speed
Together AI	Llama/Qwen/Mistral	Fast
Fireworks AI	FireAttention	Low-latency, 6 free models
Groq	LPU	>500 T/s
Cerebras	Wafer-Scale	>2000 T/s
NVIDIA NIM	91 free endpoints, DGX Cloud	20× faster than NVIDIA GPU

Hosted Open Source Models 🆓

Services that provide hosted access to open source models with API endpoints, offering easier deployment than self-hosting.

Provider	Models	Pricing	Best For
Hugging Face Inference	1000+	Free tier + $0.10-1.00/1M	Largest selection, easy deployment
Replicate	50+	Free tier + $0.20-2.00/1M	Model-as-a-service, versioning
Together AI	30+	$0.20-3.00/1M	Optimized for open models
Fireworks AI	25+	Free tier + $0.10-1.50/1M	Low-latency, 6 free models
Anyscale	100+	$0.20-2.00/1M	Distributed inference, scaling
OctoAI	20+	$0.25-2.00/1M	High-performance GPU inference
Banana.dev	15+	$0.15-1.80/1M	Fast deployment, auto-scaling
Predibase	12+	$0.30-2.50/1M	Fine-tuning + inference
RunPod Serverless	10+	$0.20-2.00/1M	GPU serverless, flexible pricing
CoreWeave	8+	Custom pricing	Enterprise-grade, dedicated GPUs

Free Tier APIs

Provider	Official free access	What happens after free or trial	Best For	Source
Google Gemini	Gemini Developer API free tier is free of charge	Gemini 2.5 Flash-Lite paid tier starts at $0.10 / $0.40	Best official always-on free multimodal API	🔗
Groq	GroqCloud Free plan is $0 for build and test use	Developer plan is pay-per-token; Llama 3.1 8B Instant starts at $0.05 / $0.08	Fast low-cost iteration	🔗, 🔗
Cohere	Trial API keys are free but limited; Command A+ is free until rate limits are reached	Command R starts at $0.15 / $0.60	RAG, tool use, multilingual testing	🔗, 🔗, 🔗
Mistral AI	Free mode is enabled by default with no credit card required	Mistral Small 4 starts at $0.15 / $0.60 on Scale	Cheap coding and multimodal experiments	🔗, 🔗
Anthropic	New users receive a small amount of free credits	Haiku 4.5 starts at $1.00 / $5.00	Testing Claude API behavior	🔗, 🔗
Vercel AI Gateway ⭐	$5/month included on the free tier for eligible models	Paid credits use provider list rates with zero markup	Multi-provider testing behind one endpoint	🔗
OpenAI	Developer quickstart includes one free test API request	Standard API pricing starts with GPT-5.4 nano at $0.20 / $1.25	SDK evaluation and first-request testing	🔗, 🔗

Automation 🤖

AI-powered tools for automating browser and desktop tasks.

Browser Automation 🌐

Tools and frameworks for AI-powered browser automation.

Standalone AI Browsers

Browser	Platform	Pricing	Open Source	Local AI	Agent/Computer Use	API Access	Multi-Agent	Parallel Sessions	Best For	GitHub
Norton Neo	Windows, macOS, iOS, Android	Free	❌	❌	⚠️	❌	❌	❌	AI-native browser, Magic Box, Peek & Summarize, Smart Tab Groups	❌
Deta Surf	macOS, Linux	Free (alpha)	✅	✅	⚠️	❌	✅	✅	Browser + file manager + AI notebook, Surf Memory	❌
Firefox (AI Controls)	Windows, macOS, Linux, iOS, Android	Free	✅	❌	⚠️	❌	❌	❌	AI Controls dashboard (v148+), ChatGPT/Claude/Mistral sidebars	❌
Dendrite	Windows, macOS, Linux	Free (OSS)	✅	✅	✅	❌	✅	✅	Developer AI browser client for agents	❌
Perplexity Comet	Windows, macOS, iOS, Android	Free / Pro $20/mo	❌	❌	✅	❌	❌	❌	Research + background tasks, voice mode, Computer Max agent	❌
ChatGPT Agent Mode	Web, iOS, Android	Plus $20/mo, Pro $200/mo	❌	❌	✅	❌	❌	❌	Full computer use: browse, code, fill forms, book travel	❌
Dia	macOS (M1+ / macOS 14+)	Free / Pro $20/mo	❌	❌	⚠️	❌	❌	❌	Tab intelligence, Skills, browsing history AI context	❌
Google Chrome (Auto Browse 2)	Windows, macOS, Linux, ChromeOS	Free / Gemini Pro $19.99/mo	❌	❌	✅	❌	❌	❌	Gemini 3 built-in, auto browse 2 agentic tasks, Chrome Skills saveable workflows, Universal Commerce Protocol, WebMCP support	❌
Microsoft Edge (Copilot Agent)	Windows, macOS, iOS, Android	Free / Copilot Pro $20/mo	❌	❌	✅	❌	❌	❌	Cross-tab context, voice commands, form automation, bookings	❌
Genspark	Windows, macOS, Web, iOS, Android	Free / Plus $25/mo / Pro $249/mo	❌	✅ (169 local models)	✅	❌	✅	✅	Super Agent, AI slides, AI websites, deep research, Call For Me, Genspark Claw desktop app	❌
Brave Leo (AI Browser)	Windows, macOS, Linux, iOS, Android	Free / Premium $14.99/mo	✅ (Chromium)	✅ (Leo local)	⚠️	❌	❌	❌	Privacy-first, zero-log AI, Skills, Memories, local models	❌
SigmaOS (Airis)	macOS	Free / Pro (subscription)	❌	❌	⚠️	❌	❌	❌	NL commands: "Book Airbnb in Iceland", cross-tab AI, YC-backed	❌
Opera Neon	Windows, macOS	$19.90/mo	❌	❌	✅	✅ MCP Connector (Mar 2026)	❌	❌	Agentic browsing, Aria assistant, MCP API for Claude/ChatGPT/n8n integration	❌
Opera One (Aria)	Windows, macOS, Linux, iOS, Android	Free	❌	❌	⚠️	❌	❌	❌	Built-in Aria AI assistant, sidebar AI tools	❌
Firefox (AI Sidebar)	Windows, macOS, Linux, iOS, Android	Free	✅	❌	⚠️	❌	❌	❌	AI Controls dashboard (v148+), ChatGPT/Claude/Mistral sidebars	❌
BrowserOS	Linux, macOS	Free	✅	✅	✅	❌	✅	✅	Privacy-focused, built-in MCP, agentic	🔗
Manus AI	Web (Cloud)	Free 300 credits/day / Plus $20/mo / Pro $200/mo	❌	❌	✅	❌	✅	✅	Cloud agent, full computer: code, deploy, files, search	❌
Sigma AI Browser	Windows, macOS, Linux	Free / Pro $29/mo	❌	✅	✅	❌	❌	❌	Built-in local AI agent, offline, no tracking	🔗
Fellou	Windows, macOS	Free 4 tasks/day / Pro $20/mo	❌	❌	✅	❌	❌	❌	Complex multi-step automation, agentic tasks	🔗
Arc Max	macOS, Windows	Free	❌	❌	⚠️	❌	❌	❌	AI-enhanced browsing, pinch-to-summarize, Ask on Page	❌
Maxthon	Windows, macOS, iOS, Android	Free / Premium	❌	❌	⚠️	❌	❌	❌	MaxAsk AI answers, built-in VPN, ad-blocker, resource sniffer	❌
ChatGPT Atlas	macOS	Free (with ChatGPT subscription)	❌	❌	✅	❌	❌	❌	OpenAI integration, macOS computer use overlay	🔗
AnythingLLM	Windows, macOS, Linux	Free (OSS)	✅	✅	⚠️	✅ (local API)	❌	❌	All-in-one desktop AI, document chat, local + API	🔗
BrowserGPT	iOS, Android	Free / Premium	❌	❌	⚠️	❌	❌	❌	Mobile-first AI browser	❌
Sidekick Browser	Windows, macOS, Linux	Free / Pro $10/mo	❌	❌	✅	❌	❌	❌	AI assistant, natural language tab management, summarize, automate tasks	❌
Browser Operator	Windows, macOS	Free (OSS)	✅	✅ (Ollama, 100+ models)	✅	❌	✅	✅	Multi-agent automation, privacy-first local processing, persistent memory	🔗
Browserless Agent ⭐	Cloud (MCP)	Free tier / paid	❌	❌	✅	✅ (MCP-native)	❌	❌	Fastest MCP browser agent; stateful sessions; command batching; released 2026-06-12	🔗
Kane CLI ⭐	Windows, macOS, Linux	Free	❌	❌	✅	❌	❌	❌	Browser automation testing for AI agents; deterministic pass/fail; released 2026-05-24	🔗
Browsewright ⭐	Python	Free (MIT)	✅	✅	✅	✅ (MCP)	❌	❌	Open-source browser agent; LLM-driven form filling; structured JSON output; released 2026-06-16	🔗
Lightpanda Agent ⭐	Windows, macOS, Linux	Free (OSS)	✅	✅	✅	❌	❌	❌	Native headless browser agent; PandaScript reproducible scripts; LLM at build-time not runtime; released 2026-06-17	🔗
Veilbrowser ⭐	TypeScript	Free (MIT)	✅	✅	✅	✅ (MCP)	❌	❌	Stealth browser for AI agents; real Chrome over raw CDP; passes sannysoft 57/57; released 2026-06-09	🔗
Dev Browser ⭐	Claude Code plugin	Free	❌	✅	✅	❌	❌	❌	Browser automation plugin for Claude Code; stateful Playwright wrapper; released 2026-06-19	🔗
Intuned ⭐	Python/Cloud	Free tier / paid	❌	✅	✅	✅	❌	❌	YC-backed browser automation as code; AI agent generates Playwright code; auto-healing; released 2026-06-08	❌
Webwright (Microsoft) ⭐	Python	Free (MIT)	✅	✅	✅	❌	❌	❌	SOTA on long-horizon web tasks; terminal-based; Codex/Claude Code plugins; released 2026-04-08	🔗
SimuLang ⭐	Python/TypeScript	Free (OSS)	✅	✅	✅	❌	❌	❌	Playwright for the entire desktop; browser + native apps + OS workflows; released 2026-05-19	🔗

Browser Extensions

Extension	Pricing	Free	Multi-Agent	Best For	GitHub
OpenDia	Free (OSS)	✅	❌	Open alternative to Dia/Comet; anti-detection for X/LinkedIn/Facebook; works with Chrome/Firefox/Edge/Brave	🔗
Monica.im	Freemium (Free + ~$9/mo)	❌	✅	Chrome extension, no browser switch	❌
Harpa AI	Free	✅	❌	Automation recipes	🔗
MultiOn	Free/Paid	⚠️	✅	Complex tasks	🔗
NanoBrowser	Free	✅	✅	Local control, Ollama	🔗
Neobrowser	Free (OSS)	✅	❌	Local LLMs via Ollama, privacy-first, Chrome/Edge	❌
Open Operator	Free	✅	❌	Browserbase-powered, open NL browser control	🔗
Openator	Free (OSS)	✅	❌	Docker-based headless NL browser agent	🔗

Developer Libraries

Consumer-focused AI browsers (for example Perplexity Comet) typically bundle chat with browsing but enforce vendor-side quotas on usage tiers. The stacks below target developers: scale via your LLM keys (BYOK where offered), metered cloud APIs, or self-hosted / OSS runtimes so automation limits follow your infrastructure and providers—not the same caps as end-user browser apps.

Library	Language	Pricing	Best For	API Access	Multi-Agent	Parallel Sessions	GitHub
Chrome DevTools MCP	TypeScript	Free (OSS)	AI web debugging, 29 DevTools	❌	❌	❌	🔗
Cloudflare Browser Run	Cloud API	Free Workers / paid browser hours	CDP + MCP, WebMCP, Live View, Human-in-the-Loop, session recordings; `/json` Quick Action extracts schema-shaped data with Workers AI or BYOK (2026-07-01)	✅	❌	✅	🔗
Browser-use ⭐	Python / Cloud API	Free OSS / Cloud free tier: 3 concurrent sessions and 10 agent tasks/mo; paid cloud from $40/mo or PAYG credits	Browser Use CLI 3.0 powered by Browser Harness; `browser-use skill` installs agent skills for Claude Code, Codex, Cursor, Gemini, OpenCode, and related skill directories; managed cloud adds stealth browsers, CAPTCHA solving, scheduling, persistent memory, skills, and integrations	✅	✅	✅	🔗
Browser Harness ⭐	Python / CDP	Free (OSS)	Thin self-healing CDP harness for direct LLM control of real Chrome; packaged skill for coding agents; pinned by Browser Use CLI 3.0	✅	❌	✅	🔗
Vercel agent-browser ⭐	Rust / Node CLI	Free (OSS)	Browser automation CLI for AI agents; compact text snapshots, ref-based actions, Chrome and Lightpanda engines, packaged agent skill	✅	❌	✅	🔗
Stagehand v3.6 ⭐	TypeScript/Python	Free (OSS)	v3.6 (2026-06-19): WebMCP support; Claude Fable 5 support; Microsoft Entra ID auth; full CDP rewrite, 44% faster, action caching, hybrid deterministic + AI	✅	❌	✅	🔗
LaVague	Python	Free (OSS)	NL to code	❌	❌	✅	🔗
Skyvern	Python	Free tier / $29–$149/mo	CV-based automation, Ollama support	✅	✅	✅	🔗
Notte	Python/Cloud	Free tier / $29/mo+	Deterministic replay, demo→script	✅	❌	✅	🔗
Firecrawl	Python / CLI	Free tier / $49/mo+	LLM-powered crawling & scraping	✅	❌	✅	🔗
Playwright MCP	TypeScript	Free (OSS)	Cross-browser automation, VS Code	✅	❌	✅	🔗
Langflow	Python	Free (OSS) / Cloud $29/mo	Visual multi-agent & RAG workflows	✅	✅	✅	🔗
LlamaIndex	Python	Free (OSS) / Cloud $29/mo	Document-heavy RAG, retrieval quality	✅	✅	✅	🔗
Haystack	Python	Free (OSS) / Cloud $49/mo	Regulated deployments, structured pipelines	✅	✅	✅	🔗
AgentQL	TypeScript/Python	Free (1K req/mo) / $49/mo / $149/mo	Natural language web querying/automation	✅	✅	✅	🔗
ScrapeGraphAI	Python	Free OSS / Cloud $29/mo	Natural language web scraping	✅	✅	✅	🔗
WebVoyager	Python	Free (OSS)	Autonomous web browsing research	❌	❌	✅	🔗
Anchor Browser	TypeScript, Python, REST	Free credits tier / usage-based plans	Managed browser fleets + AI steps for computer-use-style web tasks; see docs.anchorbrowser.io	✅	✅	✅	🔗
Hyperbrowser	TypeScript, Python, REST	Free credits + pay-as-you-go (see hyperbrowser.ai/pricing)	Hosted agents: Browser-Use, Claude/OpenAI/Gemini computer use, Stagehand, HyperAgent; optional BYOK for LLM	✅	✅	✅	🔗
Steel	Python, Node.js, REST	Hobby credits + paid tiers; OSS self-host via Docker (steel.dev, docs.steel.dev)	Sessions API for Playwright/Puppeteer/Selenium; self-hosted steel-browser avoids vendor browser caps; agent cookbooks	✅	⚠️	✅	🔗

Cloud Automation

Service	Platform	Pricing	Best For	GitHub
ChatGPT agent	ChatGPT	Plus $20/mo / Pro $200/mo / Team	Guided browser tasks, research, forms, spreadsheets, code execution	❌
Gemini Spark ⭐	Google AI Ultra	Included with Google AI Ultra ($99.99/mo, US beta from 2026-05-26)	24/7 always-on personal AI agent; runs on Google Cloud VMs while device is off; MCP connections to Canva, OpenTable, Instacart at launch; Adobe, GitHub, Notion, Slack confirmed summer 2026	❌
Claude Cowork	Anthropic (desktop app)	Included with Claude paid plans (GA 2026-04-09)	Desktop agent: works in local files & apps, multi-step task completion, Dispatch (assign by text/voice), runs while user is away	❌
Project Mariner	Google AI Ultra	Included with Google AI Ultra ($99.99/mo)	Multi-step browser tasks, shopping, and reservations	❌
Skyvern Cloud	Cloud API	Free 1K credits / Hobby $29/mo / Pro $149/mo	Resilient CV-based automation, Ollama support	🔗
Browserbase	Cloud API	Free credits / paid tiers	Stealth mode, session recording, CDP access	❌
Amazon Nova Act	API (AWS)	Pay-per-use	Autonomous web agent, multi-step browser workflows	❌
Amazon Bedrock AgentCore ⭐	AWS	Pay-per-use	Platform for building, connecting, securing, tracing, and operating production agents with any model or framework; includes Runtime, Gateway, Memory, Browser tool, Code Interpreter, Identity, and Observability	🔗
Vercel Sandbox ⭐	Vercel	Usage-based Vercel platform	Agent/runtime sandbox for code execution; FUSE mounts for S3, network, and custom filesystems added 2026-07-03; pairs with AI SDK and eve workflows	🔗
OpenAI Workspace Agents	ChatGPT Business, Enterprise, Edu, Teachers	Credit-based after preview	Research-preview workplace agents usable in ChatGPT and Slack channels; role-based admin controls	🔗
Claude Science ⭐	Anthropic	Applications through 2026-07-15 00:00 UTC	AI workbench for scientists; auditable artifacts, package/tool integration, flexible compute access, up to $30K Claude credits for selected projects	🔗
Databricks Genie One ⭐	Databricks	Free $10/user/mo	GA 2026-06-16; data-smart agentic coworker; iOS/Android/web; 50+ app integrations; Genie Ontology context layer	❌
Thoughtworks Agent/works ⭐	Multi-cloud	Enterprise	GA 2026-06-16; governed runtime for enterprise agents; single control plane; any cloud deployment	❌
Kore.ai Artemis ⭐	Azure (initial)	Enterprise	GA 2026-05-21; AI-native agent platform; governance + observability; 40+ channels; Microsoft Foundry integration	❌
Huawei Cloud AgentArts ⭐	Huawei Cloud	Enterprise	GA 2026-06-08; enterprise-grade agent platform; open-source edition (openJiuwen); AgentArts Orchard portal	❌
Zensar ZenseAI.AgentMesh ⭐	Multi-cloud	Enterprise	GA 2026-06-19; 80+ pre-built agents; governed orchestration; EU AI Act alignment	❌
Sedai AI Agent Optimization ⭐	Multi-cloud	Early access	GA planned late 2026; smart model routing; governance + observability; supports OpenAI/Bedrock/Vertex/Azure	❌
GMI Cloud AgentBox ⭐	GMI Cloud	Usage-based	GA 2026-06-08; marketplace for AI agents; 100+ models via single key; MaaS inference	❌
ClickHouse Agents ⭐	ClickHouse	Free (public beta)	GA 2026-06-09; Claude-powered agentic analytics; no-code agent builder; MCP-compatible	❌
Google Data Agents ⭐	Google Cloud	Free (preview)	GA 2026-06-16; Data Science Agent, Database Agent, Looker Dashboard Agent, Data Insights Agent, Deep Research Agent	❌
C1 Autonomous Worker ⭐	C1	Enterprise	GA 2026-06-15; agent that runs work loops, carries state, executes code; available via Slack	❌

Autonomous Agents — Plain English Prompts 🤖

Control a computer or cloud sandbox using plain English text — no coding required. Just describe the task and the agent handles everything: clicking, typing, navigating, running code, and completing multi-step workflows.

Legend: 🖥️ = runs on your physical computer | ☁️ = cloud/sandbox computer | 🌐 = controls a browser | 🔁 = multi-agent/parallel | 💬 = simple English prompt | 🔓 = free/open-source | 💰 = paid

☁️ Cloud Sandbox Computer Use (English Prompts)

These services run in a cloud sandbox (virtual Linux/Windows desktop), control the computer for you, and are driven purely by natural language instructions.

Agent	Interface	Pricing	Multi-Agent	Parallel Sessions	Local LLM	English Prompt	GitHub
Suna (Kortix)	Web dashboard	Free (OSS self-host); Cloud from $20/mo; Pro $29/mo; Team $199/mo	✅	✅	❌	✅	🔗
OpenManus	Web dashboard	Free OSS	✅	✅	❌	✅	🔗
Manus AI	Web dashboard	Free (300 credits/day) / Plus $20/mo / Pro $200/mo	✅	✅	❌	✅	❌
ChatGPT Agent	ChatGPT Web/App	Plus $20/mo / Pro $200/mo	❌	❌	❌	✅	❌
Gemini Computer Use	API / AI Studio	Gemini Pro $19.99/mo / API metered	❌	❌	❌	✅	❌
Devin	Web dashboard	Core $20/mo ($2.25/ACU) / Team $500/seat/mo	❌	❌	❌	✅	❌
OpenHands	Web UI / CLI	Free OSS / Cloud Individual free	✅	✅	✅ (any API)	✅	🔗
E2B Desktop Sandbox	API / SDK	Hobby free / Pro $150/mo	❌	✅ (via code)	✅	✅	🔗
Cua (trycua)	CLI / Python SDK	Free (OSS)	❌	✅	✅	✅	🔗
Airtop	Web dashboard / API	Starter $26/mo (3 sessions) / Pro $80/mo (30 sessions)	❌	✅	❌	✅	❌
Skyvern Cloud	Web dashboard / API	Free 1K credits / Hobby $29/mo / Pro $149/mo	❌	✅	✅ (Ollama)	✅	🔗
Convergence Proxy	Web / API	Free tier / Pro $20/mo (acquired by Salesforce)	❌	❌	❌	✅	❌
Amazon Nova Act	API (AWS)	Pay-per-use (AWS pricing)	❌	✅	❌	✅	❌
Project Mariner	Google AI Ultra	Included ($100/mo Ultra 5× plan)	❌	❌	❌	✅	❌
Perplexity Computer	Web dashboard	Perplexity Pro $20/mo	❌	❌	❌	✅	❌
OpenAI Computer Use (API)	API / ChatGPT	$15/M input, $60/M output	✅	✅	❌	✅	❌
OpenHands Agent Canvas ⭐	Web / Self-hosted	Free (OSS)	Workspace for creating automations integrating Slack/GitHub; self-hostable; released 2026-06-16	🔗

🖥️ Local Machine / Physical Computer Use

These agents run on your own machine, see your screen, and control your keyboard/mouse — no cloud required.

Agent	Windows	macOS	Linux	Dashboard/UI	CLI	API/LLM	Multi-Agent	Parallel Sessions	Pricing	GitHub
Claude Computer Use	✅	✅	✅	❌	✅ (API)	Claude API ($3–$15/M)	❌	❌	Claude API ($3–$15/M tokens)	Commercial
Agent TARS (ByteDance)	✅	✅	✅	✅ Web UI	✅ `npx @agent-tars/cli@latest`	Any LLM	❌	✅	Free (OSS)	🔗
Aiden	❌	❌	✅	✅ Desktop	✅	Local / BYO API	❌	❌	Free (OSS)	🔗
Simular Agent S2	✅	✅	✅	❌	✅	Any LLM API	❌	❌	Free (OSS)	🔗
UI-TARS Desktop (ByteDance)	✅	✅	✅	✅ Desktop app	❌	UI-TARS-2 model	❌	❌	Free (OSS)	🔗
Open Interpreter	✅	✅	✅	✅ Web	✅ `interpreter`	Any (OpenAI, Claude, local)	❌	❌	Free (OSS)	🔗
Open-Interface	✅	✅	✅	❌	✅	GPT-4V / any vision LLM	❌	❌	Free (OSS)	🔗
Agent S / S2	✅	✅	✅	❌	✅	Any LLM API	❌	❌	Free (OSS)	🔗
UFO (Microsoft)	✅	❌	❌	✅ UI	✅	GPT-4V / Azure	❌	❌	Free (OSS)	🔗
Windows-Use	✅	❌	❌	❌	✅	Any vision LLM	❌	❌	Free (OSS)	🔗
Bytebot	❌	❌	✅	✅ (Docker)	✅	Any LLM	❌	❌	Free (OSS)	🔗
OpenCUA	✅	✅	✅	❌	✅	Any	❌	❌	Free (OSS)	🔗
Windows Agent Framework (Microsoft)	✅	❌	❌	Agent Designer in VS 2026	✅ `wagent`	Any LLM	❌	❌	Free (MIT)	🔗
Khoj	✅	✅	✅	✅ Web UI	✅	Any (Ollama, LM Studio, OpenAI)	❌	❌	Free (OSS) / Cloud $10/mo	🔗
Eigent (CAMEL-AI)	✅	✅	✅	❌	✅	Any LLM	✅	✅	Free (OSS)	🔗

🌐 Browser-Only Agents

Control a browser with natural language — click, fill forms, scrape, automate. No script writing needed.

Agent	Type	Pricing	Dashboard	CLI	Multi-Agent	Parallel Sessions	Local LLM	GitHub
Browser-use ⭐	OSS Python lib + Cloud	Free OSS / Cloud free tier: 3 concurrent sessions and 10 tasks/mo; paid cloud from $40/mo or PAYG credits	✅ Cloud	✅ CLI 3.0	✅	✅	✅ (Ollama)	🔗
Stagehand	OSS TypeScript	Free (OSS)	❌	✅	❌	✅	✅	🔗
NanoBrowser	Chrome extension	Free (OSS)	✅ Extension	❌	✅	❌	✅ (Ollama)	🔗
Skyvern	Python / Cloud	Free tier / $29–$149/mo	✅ Cloud	✅	✅	✅	✅ (Ollama)	🔗
Openator	Python	Free (OSS)	❌	✅	❌	✅	✅	🔗
Open Operator	Web UI	Free	✅	❌	❌	❌	❌	🔗
Airtop	Web / API	$26–$80/mo	✅	❌	✅	✅	❌	❌
MultiOn	API / Chrome ext	Free / Paid	✅	❌	✅	❌	❌	🔗

🔁 Multi-Agent / Parallel Agent Platforms (Plain English Orchestration)

Coordinate multiple AI agents in parallel to complete complex workflows — driven by plain English goals.

Platform	Type	Dashboard	CLI	Cloud	Local LLM	Parallel	Pricing	GitHub
CrewAI	Multi-agent OSS + Cloud	✅ AMP Studio	✅ `crewai`	✅ AMP	✅	✅	Free OSS / Starter $99/mo / Pro $299/mo / Enterprise custom	🔗
Microsoft Agent Framework (AutoGen + Semantic Kernel unified)	Multi-agent conversations (v1.0 GA April 2026)	✅ AG2 Studio	✅ Python	✅ Azure	✅	✅	Free (OSS) / Azure pay-per-token	🔗
LangGraph	Stateful agent graphs	✅ LangSmith	✅ Python	✅ Cloud	✅	✅	Free OSS / Professional $99/mo	🔗
OpenHands	Dev-focused multi-agent	✅ Web UI	✅	✅ Cloud	✅	✅	Free (OSS + Cloud free tier)	🔗
OWL (Camel-AI)	Distributed multi-agent	❌	✅ Python	❌	✅	✅	Free (OSS)	🔗
Manus AI	Cloud multi-agent	✅ Web	❌	✅	❌	✅	Free 300 credits/day / $20–$200/mo	❌
n8n	Workflow + AI agents	✅ Visual canvas	✅ `n8n`	✅ Cloud	✅ (Ollama node)	✅	Free OSS / Starter $24/mo / Pro $60/mo	🔗
Devin	Software engineering	✅ Web	❌	✅	❌	❌	Core $20/mo ($2.25/ACU) / Team $500/seat	❌
Smolagents (HuggingFace)	Lightweight code agents	❌	✅ Python	❌	✅	⚠️	Free (OSS)	🔗
C3 Code (C3 AI)	Enterprise AI dev platform	✅ Web	❌	✅	❌	✅	Enterprise pricing	Full SDLC agents, governed deployment, natural language to production apps
Dify	Visual LLM platform	✅ Web UI	✅	✅ Cloud	✅	✅	Free OSS / Cloud plans	🔗
OpenAI Agents SDK	Agent handoffs + tool orchestration	❌	✅ Python / TS	✅ OpenAI API	❌	✅	Free (OSS) / OpenAI API costs	🔗
Google ADK	Hierarchical agent tree, A2A protocol	✅ Vertex AI	✅ Python	✅ Vertex AI	❌	✅	Free (OSS) / Vertex AI pay-per-token	🔗
ADK Go 2.0 ⭐	Graph-based agent workflow engine	—	✅ `go run`	✅ Google Cloud	❌	✅	Free (OSS)	🔗
Genkit Agents ⭐	Full-stack agent framework	✅ Developer UI	—	—	✅	✅	Free (OSS)	🔗
AI SDK 7 ⭐	Production agent platform (TypeScript)	—	✅ `@ai-sdk/tui`	✅ Vercel	✅	✅	Free (OSS)	🔗
Mastra ⭐	Agent framework with file-based agents, skills, workspace, and subagents	—	✅ `mastra`	✅ Vercel	✅	✅	Free (OSS)	🔗
Ruflo	Multi-agent orchestration	❌	✅ Python	❌	✅	✅	Free OSS / Cloud plans	🔗
Vercel eve ⭐	Open-source agent framework with Agent Runs traces via Vercel MCP/CLI	✅ Web	✅ `vercel agent-runs`	✅ Vercel	✅	✅	Free (OSS) / Vercel hosting costs	🔗
Omnigent (Databricks) ⭐	Meta-harness for agents	✅ Web/App	✅	✅	✅	✅	Free (Apache 2.0)	🔗
GitKraken Kepler ⭐	Agentic Development Environment	✅ Desktop	✅	❌	✅	✅	Free tier / paid plans	🔗
LangChain Deep Agents ⭐	Long-horizon agent harness	❌	✅ Python	❌	✅	✅	Free (MIT)	🔗
Wheelie (Continua) ⭐	Agentic dev OS + service runtime	✅ Web	✅	✅ VMs	✅	✅	Free (waiting list)	🔗
Cloudflare Flue ⭐	Agent framework on Workers	✅ Dashboard	✅	✅ Cloudflare	❌	✅	Free (OSS) / Workers pricing	🔗
C1 Autonomous Worker ⭐	Enterprise agent	✅ (Slack)	❌	✅	❌	❌	Enterprise	❌
ClickHouse Agents ⭐	Agentic analytics	✅	❌	✅	❌	✅	Free (public beta)	❌
Google Data Agents ⭐	Data agents	✅	❌	✅	❌	✅	Free (preview)	❌
Salesforce Agentforce 3 ⭐	MCP-native agent platform	✅ Web	❌	✅	❌	✅	Enterprise pricing	❌
Kore.ai Artemis ⭐	AI-native agent platform	✅	❌	✅ (Azure)	❌	✅	Enterprise	❌
Konecta Kolibri ⭐	Agentic AI orchestration	✅	❌	✅	❌	✅	Enterprise	❌
Zensar ZenseAI.AgentMesh ⭐	Enterprise agentic AI	✅	❌	✅	❌	✅	Enterprise	❌
Sedai AI Agent Optimization ⭐	Agent optimization	✅	❌	✅	❌	✅	Early access	❌

Multi-Agent & Parallel Execution Summary

Tools supporting parallel agent orchestration (✅) vs single-agent only (❌):

Category	Supports Parallel Agents	Tools
Cloud Sandbox	✅	Manus AI, OpenHands, E2B Desktop Sandbox, Cua (trycua), Airtop, Skyvern Cloud, Amazon Nova Act, Perplexity Computer, OpenAI Computer Use (API)
Cloud Sandbox	❌	ChatGPT Agent, Gemini Computer Use, Devin, Convergence Proxy, Project Mariner
Local Machine	✅	Agent TARS, E2B Desktop Sandbox, Cua (trycua)
Local Machine	❌	Claude Computer Use, UI-TARS Desktop, Open Interpreter, Open-Interface, Agent S/S2, UFO, Windows-Use, Bytebot, OpenCUA, Khoj
Browser-Only	✅	Browser-use, Skyvern, Airtop, MultiOn, Browser Operator
Browser-Only	❌	Stagehand, NanoBrowser, Openator, Open Operator
Developer Libraries	✅	Browser-use, Browser Harness, Vercel agent-browser, Skyvern, Cloudflare Browser Run, Langflow, LlamaIndex, Haystack, AgentQL, ScrapeGraphAI, WebVoyager, Hyperbrowser, Anchor Browser, Steel
Developer Libraries	❌	Chrome DevTools MCP, Stagehand, LaVague, Notte, Firecrawl, Playwright MCP
Multi-Agent Platforms	✅	CrewAI, Microsoft Agent Framework, LangGraph, OpenHands, OWL, Manus AI, n8n, Smolagents, Dify, OpenAI Agents SDK, Google ADK, ADK Go 2.0, Genkit Agents, AI SDK 7, Mastra, Vercel eve
Multi-Agent Platforms	❌	Devin

AI Infrastructure 🏗️

Tools, frameworks, and specialized models for building production AI systems — from embeddings and video generation to safety, evaluation, and model routing.

Embedding & Reranking Models 🧲

Specialized models for converting text (or images) into dense vector representations and for reranking retrieval results. Essential infrastructure for RAG pipelines and semantic search. Prices as of July 2026.

Embedding Models

Model	Developer	Dimensions	Max Tokens	Pricing	Best For	GitHub
Gemini Embedding 2	Google	3,072	8,192 text / 6 images / 120s video / 180s audio	$0.20/1M text tokens; GA 2026-04-22	First natively multimodal embedding — text, image, video, audio, PDF in one space	—
Granite Embedding Multilingual R2 (311M) ⭐	IBM	768	32,768	Free (Apache 2.0)	MTEB Multilingual Retrieval 65.2 (#2 open <500M); 200+ languages; 32K context (64x R1)	🔗
Granite Embedding Multilingual R2 (97M) ⭐	IBM	384	32,768	Free (Apache 2.0)	Best sub-100M multilingual embedder (MTEB 60.3); 200+ languages	🔗
pplx-embed-v1 (4B)	Perplexity	3,072	8,192	API pay-per-use	MTEB Multilingual leader; web-scale retrieval; contextual variant available	🔗
LFM2.5-Embedding-350M ⭐	Liquid AI	1,024	8,192	Free (open-source)	350M-param bidirectional embedder; 11 languages; best-in-class multilingual; runs on CPU/edge	🔗
LFM2.5-ColBERT-350M ⭐	Liquid AI	Per-token vectors	8,192	Free (open-source)	First bidirectional ColBERT; word-by-word matching for higher accuracy; 11 languages; runs on CPU/edge	🔗
ML-Embed-0.6B (CodeFuse) ⭐	CodeFuse AI	1,024	8,192	Free (Apache 2.0)	3D Matryoshka Learning; multilingual; ICML 2026 paper	🔗
Qwen3-Embedding-8B	Alibaba / Qwen	4,096 (flex)	32K	Free (Apache 2.0)	MTEB leader (70.6); 100+ languages, instruction-aware, flexible dimensions; 0.6B/4B/8B sizes	🔗
jina-embeddings-v5-omni-small ⭐	Jina AI / Elastic	1,024 (Matryoshka 32-1,024)	32K text; image, video, audio, PDF	API pay-per-use	Multimodal embedding model for text, image, video, audio, and PDF search; frozen v5 text backbone keeps text outputs aligned with v5-text-small	🔗
jina-embeddings-v4	Jina AI	32–2,048 (flex)	8,192	API pay-per-use	Built on Qwen2.5-VL-3B (3.8B); 3 LoRA adapters (query/passage/matching)	🔗
text-embedding-3-small	OpenAI	1,536	8,191	$0.02/1M tokens	Cost-effective English embeddings	—
text-embedding-3-large	OpenAI	3,072	8,191	$0.13/1M tokens	Highest-quality English retrieval	—
Embed v4	Cohere	1,536	128K	$0.12/1M (text), $0.47/1M (image)	Multimodal text + image RAG	—
voyage-3-large	Voyage AI	256–2,048 (flex)	32K	~$0.18/1M tokens	Highest-quality retrieval, long context	—
BGE-M3	BAAI	1,024	8,192	Free (open-source)	Multi-functional: dense + sparse + ColBERT	🔗
Nomic Embed v2 (MoE)	Nomic AI	256–768 (flex)	512	Free (open-source)	Multilingual, MoE efficiency (305M active)	🔗
text-embedding-005	Google (Vertex AI)	768	2,048	$0.10/1M tokens	GCP-native semantic search	—

Reranking Models

Model	Developer	Max Tokens	Pricing	Best For	GitHub
Rerank 4.0 Pro	Cohere	32K	$1.00/1K queries	High-accuracy domain-specific reranking	—
Rerank 4.0 Fast	Cohere	32K	$0.50/1K queries	Low-latency production reranking	—
rerank-2.5	Voyage AI	32K	API pay-per-use	Instruction-following, multilingual	—
Qwen3-Reranker-4B	Alibaba / Qwen	32K	Free (Apache 2.0)	MTEB-R 69.76, MMTEB-R 72.74; 100+ languages; instruction-following	🔗
BGE Reranker v2-m3	BAAI	8,192	Free (open-source)	Open-source cross-encoder reranking	🔗
Jina Reranker v2	Jina AI	8,192	API pay-per-use	Multilingual, long-context reranking	—

Video Generation Models 🎬

Text-to-video and image-to-video generation models for creating short clips from prompts. The field is moving rapidly — resolutions, durations, and pricing change frequently. Specs as of July 2026.

Model	Developer	Resolution	Duration	Pricing	Open Source	Best For	GitHub
~~Sora 2~~ (deprecated)	OpenAI	Up to 1080p	Up to 20s	App shut down 2026-04-26; API shutdown 2026-09-24	No	Deprecated — use Veo 3 or Luma Ray3	—
Veo 3.1	Google DeepMind	1080p	Up to 8s (extendable)	~$0.20–$0.40/s	No	Native 48kHz audio + video, realistic physics; only model with synced dialogue	—
Gemini Omni Flash ⭐	Google DeepMind	1080p	Public preview	$0.10/s	No	High-quality video generation + conversational editing; text/image/video input; released 2026-06-30	—
Seedance 2.0	ByteDance	1080p	Up to 16s	API pay-per-use	No	#1 Artificial Analysis leaderboard (Elo 1213, June 2026); accepts 9 images + 3 clips + 3 audio	—
Luma Ray3.14	Luma AI	Up to 1080p	Up to 10s	$29–$99/mo	No	Native 1080p, 4× faster at 720p vs Ray2, 3× lower cost	—
LTX-2	Lightricks	Up to 4K	Up to 20s @ 50fps	Free (open-source; commercial license for >$10M ARR)	Yes (Apache 2.0)	First open-source 4K audio+video sync model, 19B DiT params, runs on consumer GPU	🔗
Runway Gen-4 / Gen-4.5	Runway	Up to 4K	Up to 16s	$12–$76/mo	No	Professional creative workflows	—
Kling 3.0	Kuaishou	1080p	Up to 15s	Free / $5.99–$66/mo	No	Multi-Shot Storyboard, best value-for-money; released 2026-02-04	—
Pika 2.0	Pika Labs	1080p	Up to 5s	Free / $8–$58/mo	No	Social media, creative effects	—
MiniMax Video-01	MiniMax	720p	Up to 6s	~$0.40/video	No	Strong text-motion responsiveness	—
HunyuanVideo	Tencent	720p–2K	Up to 16s	Free (self-host; ~60GB VRAM)	Yes (Apache 2.0)	High per-frame fidelity, long clips	🔗
Wan 2.2 (14B)	Alibaba	480p–1080p	Up to 10s	~$0.10–$0.30/clip (API)	Yes (Apache 2.0)	Motion quality, VBench #1 benchmark	🔗
MiniMax Hailuo 2.3 ⭐	MiniMax	1080p	Up to 10s	API pay-per-use	No	Released 2026-06-22; successor to Hailuo 02	—
MiniMax Hailuo 02 ⭐	MiniMax	1080p	Up to 10s	API pay-per-use	No	3x params vs predecessor, 4x training data, best complex instruction adherence; #2 Artificial Analysis Video Arena; released 2026-06-20	—
Kling 3.0 Turbo ⭐	Kuaishou	480p–720p	1–15s preview	Free / paid	No	Fast-preview mode for rapid iteration; released 2026-06-17	—
MiniMax Video-01 ⭐	MiniMax	720p	Up to 6s	~$0.40/video	No	First AI-native video model from MiniMax; text-to-video + image-to-video; released 2026-06-20	—
Grok Imagine Video 1.5 ⭐	xAI	720p	Up to 15s	$0.06/s ($4.20/min at 720p)	No	#1 Image-to-Video Arena leaderboard (Elo 1,473); native audio+video sync; Fast variant 25s for 6s clip; GA 2026-06-16	—
HappyHorse-1.0 ⭐	Alibaba (ATH)	1080p	Up to 5s	Free (open-source)	Yes (open-source)	#1 on Artificial Analysis video leaderboard; unified Transformer; simultaneous audio/video; 15B params; API coming 2026-04-30	🔗
Seedance 2.1 ⭐	ByteDance	1080p	Up to 16s	API pay-per-use	No	Sharper motion, tighter character consistency; available on Atlas Cloud Day 0	—
Microsoft Mirage ⭐	Microsoft Research	—	—	Free (research)	Research	Video world model with persistent spatial memory; 10.57x faster generation; built on Wan2.2	🔗
MoVerse ⭐	Academic/MSFT	—	—	Free (research)	Research	Real-time video world model; panoramic Gaussian scaffold; 8 FPS on RTX 4090	—
Echo-Infinity ⭐	JD/Echo Team	—	24h+ rollouts	Free (research)	Research	Real-time infinite video generation; learnable evolving memory; 1.3M+ frames	🔗
JoyAI-Echo ⭐	JD	Up to 1080p	Minute-level	Free (research)	Research	Multi-shot audio-video generation; 7.5x speedup via DMD distillation	🔗
Mochi 1	Genmo	480p	Up to 5.4s @ 30fps	Free (open-source)	Yes (Apache 2.0)	High-quality open text-to-video	🔗
CogVideoX	Zhipu AI / Tsinghua	720p	~6s	Free (open-source)	Yes (Apache 2.0)	Image-to-video quality, LoRA fine-tuning	🔗

Speech & TTS Models 🔊

Text-to-speech (TTS) and speech-to-text (STT / ASR) models for voice generation, transcription, and real-time audio. Prices as of July 2026.

Artificial Analysis Speech Arena — Top 5 (July 2026)

Rank	Model	Elo	Pricing	Languages
🥇 #1	Sonic-3.5 ⭐	~1,220	~$39 / 1M chars	40+
🥈 #2	Gemini 3.1 Flash TTS	~1,216	$18.30 / 1M chars	70+
🥉 #3	Realtime TTS-2 (Research Preview)	~1,208	$25–$35 / 1M chars	100+
#4	Sonic 4	~1,210	~$46.70 / 1M chars	40+
#5	Realtime TTS 1.5 Max	~1,200	$35 / 1M chars	100+

Arena Elo scores shift continuously. Treat rankings as point-in-time readings. Source: Artificial Analysis, June 2026. Cartesia positions Sonic-3.5 and Ink-2 as its latest production speech stack for real-time voice agents. Source: 🔗

Text-to-Speech (TTS) — Proprietary & API

Model	Developer	Languages	Real-time	Open Source	Pricing	Best For	GitHub
Sonic-3.5 ⭐	Cartesia	42	Yes (sub-90ms)	No	See Cartesia pricing	Fastest, most natural Cartesia TTS model; ranked #1 for naturalness; GA in 2026 changelog	🔗
Gemini 3.1 Flash TTS	Google	70+	Yes	No	$18.30 / 1M chars	Audio tags for granular style/pace control, SynthID watermarking	—
Realtime TTS-2 (Research Preview)	Inworld AI	100+	Yes	No	$25–$35 / 1M chars	Realtime conversation, cross-lingual voice identity, emotional perception	—
Sonic 4 ⭐	Cartesia	40+	Yes	No	~$46.70 / 1M chars	Sonic 4 Turbo ~40ms TTFA (May 2026); 40+ languages, 95% world pop; instant 3s voice clone	—
Realtime TTS 1.5 Max	Inworld AI	100+	Yes	No	$35 / 1M chars	Realtime conversational agents, low latency + low cost	—
Eleven v3	ElevenLabs	70+	No	No	Subscription / API (up to 55% price cut May 2026)	Most expressive TTS, inline audio tags `[whispers]`, `[laughs]`, multi-speaker dialogue	—
Eleven Flash v2.5	ElevenLabs	32	Yes (~75ms)	No	Subscription / API	Ultra-fast real-time, same voice library as offline	—
Eleven Multilingual v2	ElevenLabs	30	Yes	No	Subscription / API	Emotionally-aware multilingual synthesis	—
gpt-4o-mini-tts	OpenAI	50+	Yes	No	$0.60 / $12 per 1M (in/out)	Natural-language voice steering, 13 built-in voices	—
GPT-Realtime-2	OpenAI	—	Yes	No	—	Speech-to-speech with GPT-5-class reasoning, tool calls, interruptions — released 2026-05-07	—
GPT-Realtime 2.1 / 2.1-mini ⭐	OpenAI	—	Yes	No	—	Newer Realtime API speech-to-speech models with function calling, MCP servers, and SIP; shipped 2026-07-06	🔗
gpt-realtime-mini	OpenAI	—	Yes	No	—	Cost-efficient speech-to-speech	—
GPT-Live-1 / 1 mini ⭐	OpenAI	—	Yes (full-duplex)	No	Bundled in ChatGPT tiers (API pending)	Consumer voice model powering ChatGPT Voice; full-duplex, listens while speaking; GPT-5.5 backend delegates complex work; announced 2026-07-08, API "soon"	🔗
gpt-audio-mini	OpenAI	—	Yes	No	—	Cost-efficient audio generation	—
OpenAI TTS / TTS HD	OpenAI	57	Yes	No	$15 / $30 per 1M chars	Enterprise, seamless GPT integration	—
Grok TTS	xAI	25+	Yes	No	$15.00 / 1M chars	Speech tags, 80+ voices, SOC 2 / HIPAA compliant	—
Deepgram Aura-2	Deepgram	7	Yes (~90–200ms)	No	$0.030 / 1K chars ($30/1M)	Enterprise voice agents, unified STT+TTS stack, on-prem deployment	—
Hume Octave 2	Hume AI	11+	Yes (~100–200ms)	No	Varies (contact sales)	Emotional intelligence, voice conversion, phoneme editing	—
Supertonic 3 ⭐	Supertone	31	Yes	No	Subscription / API	Fast, cost-efficient multilingual narration	—
Lightning V3.1 / V3.2	Smallest.ai	15	Yes	No	Pay-as-you-go	Conversational TTS, MOS 3.89, auto language detection, mid-sentence switching	—
StepAudio 2.5 Realtime ⭐	StepFun	Chinese, English	Yes	No	—	End-to-end real-time speech LLM, paralinguistic comprehension, persona RLHF	—
VibeVoice-Realtime-0.5B ⭐	Microsoft	English (9 experimental langs)	Yes (~300ms)	Yes (MIT)	Free (self-host)	0.5B real-time TTS; streaming text input; 10-min robust generation; Qwen2.5 base; <300ms TTFA; released 2025-12-03; HuggingFace integration 2026-03-06	🔗
Qwen3-TTS 🇨🇳	Alibaba	10	Yes (streaming)	Yes (Apache 2.0)	Free (self-host) / API	Voice design, voice cloning, instruction control, multilingual	🔗
Stability Audio 3.0	Stability AI	Music/SFX	Yes (small/medium OSS)	Yes (small/medium, Apache 2.0)	Free (small/medium OSS weights); commercial via API	Professional-grade music >6 min, open-weight small/medium variants	🔗
Sesame CSM	Sesame AI Labs	English	Yes	Yes	Free	Conversational, emotionally expressive (4.7 MOS)	🔗
MAI-Voice-2 ⭐	Microsoft	15	Yes	No	API (Foundry)	Most expressive Microsoft TTS; granular emotion control (whisper/sad/etc.); 15 languages; custom voice from 5-60s clip; released 2026-06-02	—
Higgs Audio v3 TTS ⭐	Boson AI	100+	Yes	No	API / free (self-host)	Voice chat optimized; inline emotion/style/prosody tags; zero-shot voice cloning; released 2026-06-04	🔗
Chatterbox Multilingual v3 ⭐	Resemble AI	25	Yes	Yes (MIT)	Free (self-host) / API	0.5B Llama backbone; PerTh watermarking; improved speaker similarity; released 2026-06-10	🔗
MisoTTS ⭐	Miso Labs	—	Yes	Yes (Modified MIT)	Free (self-host) / API pending	8B emotive TTS; RVQ scales vocabulary; conditions on text + audio; released 2026-06-03	🔗
MiniMax Speech 2.8	MiniMax	40+	Yes	No	API pay-per-use	Native sound tags, high-fidelity cloning, studio-grade clarity; introduced 2026-01-23 00:00 UTC	🔗
ZONOS2 ⭐	Zyphra	Multilingual	Yes	Yes (Apache 2.0)	Free (self-host) / API	8B MoE TTS; first open-source MoE TTS; multilingual + code-switched; zero-shot voice cloning; ECAPA-TDNN speaker embeddings; released 2026-06-12	🔗
Mistral Voxtral TTS ⭐	Mistral AI	9	Yes	Yes (CC BY-NC 4.0)	$0.016/1K chars	4B parameter TTS; 70ms latency; voice cloning from 3s audio; released 2026-06-18	🔗
MiniMax Speech-01-HD ⭐	MiniMax	17	Yes	No	API	300+ pre-built voices; high-fidelity voice cloning from 10s audio; released 2026-06-19	—
dots.tts ⭐	RedNote (Xiaohongshu)	—	Yes	Yes (Apache 2.0)	Free	2B fully continuous AR TTS; 48kHz AudioVAE; no discrete tokens; released 2026-06	🔗

Speech-to-Text (STT / ASR)

Model	Developer	Languages	Real-time	Open Source	Pricing	Best For	GitHub
Ink-2 ⭐	Cartesia	40+	Yes (~100ms)	No	Credit-based	Real-time transcription model paired with Sonic-3.5 for voice-agent workflows; launched 2026-06-30 00:00 UTC	🔗
MAI-Transcribe-1.5 ⭐	Microsoft	43	Yes	No	API (Foundry)	SOTA on FLEURS; #3 Artificial Analysis; 5x faster than Gemini 3.1; keyword biasing; 1hr audio in <15s; released 2026-06-02	—
Cohere Transcribe ⭐	Cohere	14	No	Yes (Apache 2.0)	Free (self-host) / Model Vault	#1 HuggingFace Open ASR Leaderboard (5.42% WER); 2B Conformer; released 2026-03-26
VibeVoice-ASR-7B ⭐	Microsoft	50+	Yes (~15s for 60min audio)	Yes (MIT)	Free (self-host)	60-min single-pass ASR; speaker diarization + timestamps; 50+ languages; structured Who/When/What output; released 2026-01-21; HuggingFace transformers integration 2026-03-06	🔗
Cohere Transcribe v2 ⭐	Cohere	14	Yes	Yes (Apache 2.0)	Free (self-host)	Real-time streaming STT; open-source; released 2026-05	🔗
Gladia Solaria-3 ⭐	Gladia	5 (EU)	Yes	No	API	#1 on business audio; optimized for noisy/real-world European languages; released 2026-06-10	—
Speechmatics Melia ⭐	Speechmatics	55+	Yes (preview)	No	From $0.129/hr (10hr free)	Code-switching across 55+ languages; lowest-priced Speechmatics model; released 2026-06-17	—
Voxtral Realtime ⭐	Mistral AI	13	Yes	Yes (Apache 2.0)	Free (self-host)	Natively streaming ASR; 480ms delay matches Whisper quality; 4B params; released 2026-02	🔗
Soniox v5 Async ⭐	Soniox	—	No	No	API	Structured speech-to-text; speaker separation; language IDs; normalized structured entities; released 2026-06-11	—
NVIDIA Nemotron 3.5 ASR Streaming 0.6B ⭐	NVIDIA	40	Yes	Yes (Apache 2.0)	Free (self-host) / API	40 language-locales; language-ID prompt conditioning; punctuation/capitalization; released 2026-06-04	🔗
Gnani Prisma v2.5 ⭐	Gnani AI	9 (Indian)	Yes	No	API	#1 in 8/9 Indian languages on real-world benchmarks; 15% lower WER for rural Hindi; 14M hours training data; released 2026-06-19	—
Grok STT	xAI	25+	Yes	No	$0.10/hr (batch) / $0.20/hr (streaming)	Entity recognition (medical/legal/financial), diarization, multichannel, SOC 2 / HIPAA	—
Whisper large-v3	OpenAI	100+	No	Yes (MIT)	$0.006/min (API)	Open-source multilingual baseline	🔗
GPT-4o Transcribe	OpenAI	50+	Yes	No	$0.006/min	High-accuracy managed STT	—
GPT-4o-mini-transcribe	OpenAI	50+	Yes	No	—	Cost-efficient managed STT	—
Deepgram Nova-3	Deepgram	36+	Yes	No	$0.0043/min	Ultra-low latency, production STT	—
Deepgram Nova-3 Fast	Deepgram	36+	Yes	No	—	Lowest-latency Deepgram tier	—
AssemblyAI Universal-2	AssemblyAI	Multilingual	Yes	No	$0.0025/min	Accurate, feature-rich transcription	—
AssemblyAI Universal-3 Pro ⭐	AssemblyAI	Multilingual	Yes	No	API (pay-per-use)	New flagship STT; Universal-3 Pro + Multilingual Streaming (May 2026); Voice Agent product; LLM Gateway	—
Inworld AI Multilingual Streaming ⭐	Inworld AI	Multilingual	Yes	No	API	Real-time multilingual STT for voice agents; launched May 7, 2026 alongside Universal-3 Pro	—
ElevenLabs Scribe v2	ElevenLabs	90+	Yes	No	—	State-of-the-art transcription, word-level timestamps, diarization	—
ElevenLabs Scribe v2 Realtime	ElevenLabs	90+	Yes (~150ms)	No	—	Live transcription, ultra-low latency	—

TTS Pricing Comparison (Per 1M Characters)

Provider	Model	Cost / 1M chars	Languages	Voice Cloning
VibeVoice-Realtime	vibevoice-realtime-0.5b	Free	English (9 experimental)	✅
Mistral Voxtral TTS	voxtral-tts	$16	9	✅
Fish Audio S2 Pro	s2-pro	$15	80+	✅
OpenAI	tts-1	$15	57	❌
Grok TTS	grok-tts	$15	25+	✅
Gemini 3.1 Flash TTS	gemini-3-1-flash-tts	$18.30	70+	❌
Cartesia	sonic-3.5	~$39	40+	✅ (instant)
Deepgram Aura-2	aura-2	$30	7	❌
Inworld	realtime-tts-1.5-max	$35	100+	✅
ElevenLabs	flash-v2.5	~$60 (effective)	32	✅
MiniMax	speech-2.8-turbo	$60	40+	✅
OpenAI	tts-1-hd	$30	57	❌

Use-Case Recommendations

Use Case	Top Picks
Real-time voice agents	Cartesia Sonic-3.5 (~100ms TTFA), Ink-2 (~100ms STT), Inworld Realtime TTS-2, Deepgram Aura-2 (~90ms)
Long-form narration / audiobooks	ElevenLabs v3, Gemini 3.1 Flash TTS, Fish Audio S2 Pro
Multilingual content	Gemini 3.1 Flash TTS (70+), ElevenLabs v3 (70+), Fish Audio S2 Pro (80+), Sonic-3.5 (40+)
Emotional fidelity	Hume Octave 2 (reads for meaning), ElevenLabs v3 (audio tags), Fish Audio S2 Pro
On-device / low cost	Kokoro-82M, MOSS-TTS-Nano, Qwen3-TTS (0.6B), Voxtral TTS (4B), VibeVoice-Realtime (0.5B)
Video dubbing (duration control)	IndexTTS-2 (precise duration control + emotion)
Enterprise / compliance	Deepgram Aura-2 (on-prem, HIPAA), Grok TTS (SOC 2 / HIPAA)
Open-source production	Fish Audio S2 Pro, MOSS-TTS-v1.5, IndexTTS-2, Chatterbox Multilingual v3, ZONOS2

AI Safety & Guardrails 🛡️

Tools and frameworks for detecting unsafe content, preventing prompt injection, validating outputs, and enforcing policy compliance in LLM-powered applications. As of July 2026.

Tool	Developer	Type	Open Source	Pricing	Best For	GitHub
Llama Guard 3	Meta	Safety classifier (8B LLM)	Yes (Meta license)	Free / ~$0.02/1M tokens (API)	Input/output safety classification, 8 languages	🔗
NeMo Guardrails	NVIDIA	Programmable guardrail toolkit (Colang DSL)	Yes (Apache 2.0)	Free	Dialog safety, policy enforcement, LangChain-native	🔗
OpenAI Privacy Filter	OpenAI	PII detection & redaction	Yes (Apache 2.0)	Free (OSS)	Detects & redacts personal info in text	🔗
Guardrails AI	Guardrails AI	Python validator framework	Yes	Free (OSS)	Output validation, PII detection, hallucination guards	🔗
Amazon Bedrock Guardrails	AWS	Managed safety layer	No	Pay-per-use (AWS)	AWS-native, zero-ops compliance and content filtering	—
ShieldGemma 2	Google	Safety classifier (open weights)	Yes (open weights)	Free	Text safety (2B/9B/27B), image safety (4B)	—
LLM Guard	Protect AI	Open-source middleware toolkit	Yes (MIT)	Free	PII + toxicity filtering; chains multiple scanners; drop-in middleware	🔗
Rebuff	Protect AI	Prompt injection detector	Yes	Free	Self-hardening anti-injection using vector memory	🔗
Lakera Guard	Lakera	Managed LLM security API	No	Free tier + Enterprise	Runtime LLM security, <50ms latency, PII + injection	—
Sponsio ⭐	Sponsio Labs	Runtime contract enforcement	Yes (Apache 2.0)	Free (OSS)	Deterministic agent safety; trajectory-level policy enforcement; <0.01ms per tool call; OWASP Agentic Top 10 bundles; released 2026-05-06	🔗
ToolShield ⭐	CHATS-lab (ICML 2026)	Multi-turn safety defense	Yes (MIT)	Free	Training-free MCP tool guard; 30% attack reduction; plug-and-play for 5 coding agents	🔗
AgentGuard ⭐	WhitzardAgent	Attribute-based access control	Yes (GPL-3.0)	Free	Tool-call-level access control; supports LangChain/AutoGen/OpenAI Agents SDK	🔗
Mirage ⭐	ysham123	Policy gateway	Yes (MIT)	Free	Deterministic policy DSL; same YAML for CI and production; no LLM in decision loop	🔗
ToolSafe ⭐	MurrayTom	Step-level guardrail	Yes	Free	Proactive monitoring + feedback-driven reasoning; reduces harmful tool executions	🔗
SafeHarbor ⭐	ljj-cyber (ICML 2026)	Memory-augmented guardrail	Yes	Free	Hierarchical Risk Tree; Safety Projector; no fine-tuning of underlying model	🔗
NEXUS ⭐	eliashossain001	Runtime safety monitor	Yes	Free	Structured plan IR; 4 graded interventions (allow/block/confirm/revise); deterministic + learned	🔗
mech-gov-framework ⭐	Santander AI Lab	Mechanical governance	Yes (Apache 2.0)	Free	Model-agnostic governance regimes (R1/R2/R3); hard gates; entropy commit-reveal	🔗
Vermillio ⭐	Vermillio	Guardrails-as-a-Service SDK	No	Paid	Real-time prompt/response filtering; copyright and data leakage protection; modular safety filters	—
Amazon Bedrock Guardrails InvokeGuardrailChecks API ⭐	AWS	Per-request safeguard checks	No	Pay-per-use	Granular per-turn safety checks; numeric scores; custom thresholds; detect-only mode; released 2026-06-16	—
Google AI Control Roadmap ⭐	Google DeepMind	Agent security framework	No (internal)	Free (published)	Defense-in-depth; MITRE ATT&CK-based threat modeling; AI supervisor monitoring; D1-D4/R1-R3 levels; published 2026-06-18	—

RAG Frameworks 🗂️

Frameworks and libraries for building Retrieval-Augmented Generation (RAG) pipelines — connecting LLMs to external knowledge sources. As of July 2026.

Framework	Developer	Language	Key Features	Open Source	GitHub
LlamaIndex	LlamaIndex	Python	160+ data connectors, hybrid search, multi-agent support	Yes (MIT)	🔗
LangChain	LangChain AI	Python / JS	Chains, agents, memory, 50K+ integrations, LangGraph	Yes (MIT)	🔗
Dify ⭐	LangGenius	Python / JS	131K+ GitHub stars; visual workflow builder, RAG pipelines, MCP client/server, multi-agent, Human Input node, 1M+ apps deployed; $30M Series Pre-A (Mar 2026)	Yes (Apache 2.0)	🔗
RAGFlow	InfiniFlow	Python	Visual workflow builder, deep document parsing (PDF/tables)	Yes (Apache 2.0)	🔗
Haystack	deepset	Python	Modular pipelines, enterprise-grade, built-in monitoring	Yes (Apache 2.0)	🔗
Verba	Weaviate	Python	No-code UI, Weaviate-native vector search	Yes	🔗
Mem0	Mem0 AI	Python / JS	Persistent memory layer, graph memory, session recall	Yes (Apache 2.0)	🔗
txtai	NeuML	Python	All-in-one semantic search + workflow automation	Yes (Apache 2.0)	🔗
R2R	SciPhi	Python	Lightweight, low-latency, REST API, production-first	Yes (MIT)	🔗
Ultra RAG ⭐	rblake2320	Python	7-stage ingestion + 10-step query pipeline; KG+PPR, RAPTOR, CRAG, Self-RAG, HyDE; adversarial self-testing	Yes (MIT)	🔗
ForgeRAG ⭐	deeplethe	Python	Production-ready; structure-aware reasoning; BM25+vector+KG; pixel-precise citations	Yes (MIT)	🔗
GRIP (ACL 2026) ⭐	WisdomShell	Python	Retrieval-as-generation; token-level retrieval control; self-triggered information planning	Yes (MIT)	🔗
MOTHRAG ⭐	Julian Geymonat	Python	Deterministic multi-hop RAG; research-SOTA parity on commodity APIs; no GPU; proof tree per answer	Yes (Apache 2.0)	🔗
AkasicDB / Omni RAG ⭐	KAIST/GraphAI	SQL/GQL	Unified vector-graph-relational DBMS; Omni RAG improves accuracy 78% vs conventional RAG; 20x faster queries	Yes (SIGMOD 2026)	—
Ennoia ⭐	vunone	Python	Declarative Document Indexing; typed schemas; hybrid filter+vector search; MCP tool surface	Yes (Apache 2.0)	🔗
VORTEXRAG ⭐	vignesh2027	Python	7-layer pipeline solving semantic drift + context poisoning; EM=74.8 on multi-hop QA	Yes (MIT)	🔗
MemGraphRAG ⭐	XMUDeepLIT	Python	Memory-based multi-agent graph RAG; three-layer memory architecture; KDD 2026	Yes (MIT)	🔗
UnWeaver ⭐	Academic	Python	Entity-based decomposition RAG; better than GraphRAG precision without explicit graph; ICLR 2026	Yes	🔗
Google Agentic RAG ⭐	Google	Cloud	Multi-agent workflow with sufficient context agent; +34% accuracy vs standard RAG; public preview 2026-06-05	No	—

Fine-tuning Platforms ⚙️

Tools and platforms for adapting pre-trained LLMs to specific tasks or domains via supervised fine-tuning, RLHF, LoRA/QLoRA, and related methods. Prices as of July 2026.

Platform	Type	Supported Models	Pricing	Best For	GitHub
Unsloth	OSS library	Llama, Mistral, Gemma, Qwen, Phi, DeepSeek, GLM, + more	Free	2–5× faster training, 80% VRAM reduction; MoE 12× faster (2026), FP8 RL support (1.4× faster, 60% less VRAM); Unsloth Studio web UI; Windows officially supported	🔗
Axolotl	OSS framework	Most Hugging Face models	Free	Config-as-code (YAML), reproducibility, multi-GPU training	🔗
OpenAI Fine-tuning	Managed API	GPT-4.1, GPT-4.1-mini (SFT/DPO), o4-mini (RFT)	GPT-4.1: ~$3.00/1M training tokens; GPT-4.1-mini: ~$0.80/1M	Managed, no infra; note: closed to new users as of 2026 — existing users only	—
Google Vertex AI	Managed cloud	Gemini 3.1 Pro/Flash, Gemma 4	Gemini 3.1 Pro: $25/1M training tokens	GCP-native, Gemini model access	—
SiliconFlow	Managed cloud	100+ open-source models	Free tier + pay-per-use	Managed fine-tuning + inference; 3-step pipeline (upload→train→deploy); 2.3× faster than avg cloud; H100/H200/MI300	🔗
Unsloth Studio ⭐	Unsloth	Desktop app	500+ models	Free (OSS) / Pro / Enterprise	No-code local training and inference; runs 100% offline; GGUF/Safetensors; tool-calling; web search; released 2026-06-18
Arkor ⭐	WlyZhang	TypeScript	Open-weight LLMs	Free (alpha)	TypeScript framework for fine-tuning; type-safe configs; local Studio; managed GPUs
Langtrain ⭐	Langtrain	Python	20+ open-source models	Free (public beta)	Sovereign AI platform; fine-tune, align, deploy on own infrastructure; no per-token cost
Tuning Engines ⭐	ShinyLaunch	Python	100+ models	Free tier + paid	Unified AI control and governance layer; single OpenAI-compatible endpoint; fine-tuning + inference + guardrails
Predibase / LoRAX	Cloud + OSS server	Llama, Mistral, 50+ HF models	Free tier + per-GPU pricing	Multi-adapter serving: many LoRA adapters on one GPU	🔗
PEFT	Hugging Face	All Hugging Face models	Free	LoRA, QLoRA, prefix tuning, prompt tuning — full HF ecosystem	🔗
LLaMA-Factory	Community	100+ models	Free	Web UI, low-code interface, beginner-friendly fine-tuning	🔗
torchtune	PyTorch	Llama, Gemma, Mistral, Phi	Free	PyTorch-native, composable training recipes	🔗
Pioneer (Fastino Labs) ⭐	Managed agentic fine-tuning	Qwen, Gemma, Llama, GLiNER	API-based	First agentic fine-tuning agent; synthetic dataset generation; 10-min fine-tuning; adaptive inference	🔗
Langtrain ⭐	Managed platform	20+ open-source models	Public beta (free)	Sovereign AI platform; LoRA/QLoRA/DPO/GRPO; 100% on-prem; SOC 2 Type II	🔗
Fireworks Training ⭐	Managed cloud	100B+ models (Kimi K2.5 1T, etc.)	Preview (contact sales)	Full-parameter training; custom loss functions; frontier RL; multi-LoRA serving	🔗
Together AI Fine-tuning ⭐	Managed cloud	100B+ open-source models	Pay-per-use	Multi-node orchestration; 100B+ models; fine-tuning + inference in one platform	🔗
Tuning Engines ⭐	Unified AI control layer	100+ models	Free tier + paid	Single OpenAI-compatible endpoint; fine-tuning + routing + guardrails + policy-as-code	🔗
NeuralForge ⭐	Local-first platform	1-3B models on consumer GPUs	Free (OSS)	QLoRA training on consumer GPUs; web UI; GGUF export; single 12GB card sufficient	🔗
LLM Fine-Tuner v3.2 ⭐	No-code local tool	Most HF models	Free (GPL-3.0)	No-code web UI; SFT/DPO/RLHF/ORPO; GGUF export; Unsloth 2-5x acceleration	🔗

Evaluation & Observability 📊

Tools for tracing LLM calls, evaluating output quality, debugging RAG pipelines, and monitoring production AI systems. Prices as of July 2026.

Tool	Developer	Type	Open Source	Pricing	Best For	GitHub
LangSmith	LangChain AI	Tracing + evaluation platform	No (enterprise self-host)	Free (5K traces/mo), paid plans	LangChain apps, chain + agent debugging	—
Braintrust	Braintrust Data	Eval-first platform	Partial (AI proxy OSS)	Free (1M spans), enterprise	CI/CD evals, dataset management, LLM-as-judge; Topics active observability (production trace pattern discovery) GA June 2026	—
Helicone	Helicone	Proxy-based observability	Yes	Free tier, usage-based	Cost tracking, request caching, drop-in API proxy	🔗
Arize Phoenix	Arize AI	OSS tracing + evaluation	Yes	Free (OSS); Arize Cloud paid	RAG debugging, LLM-as-judge, local dev	🔗
Langfuse	Langfuse	Tracing + evaluation	Yes (MIT)	Free / self-host; cloud paid	Open-source, 19K+ GitHub stars, OpenTelemetry	🔗
MLflow	Linux Foundation / Databricks	Full AI engineering platform	Yes (Apache 2.0)	Free (OSS); Databricks managed paid	30M+ monthly downloads; observability, eval, prompt optimization, governance — no enterprise paywall	🔗
Ragas	Ragas	RAG evaluation framework	Yes	Free	RAG-specific metrics: faithfulness, recall, precision	🔗
DeepEval	Confident AI	LLM evaluation framework	Yes	Free (OSS); cloud paid	14+ built-in metrics, pytest-style eval runner, 50+ research-backed metrics, production anomaly detection	🔗
Laminar ⭐	Open-source observability	Yes (MIT)	Free (OSS); managed platform paid	Rust-based; ultra-fast; OpenTelemetry-native; traces + evals + AI monitoring; YC S24	🔗
Peekr ⭐	Zero-config observability	Yes (MIT)	Free	Auto-instruments OpenAI/Anthropic/LiteLLM; claim-level hallucination detection; HIPAA/GDPR guardrails	🔗
TraceMind ⭐	Open-source eval + observability	Yes	Free	Self-hosted; automatic quality scoring; eval suites; regression alerts; hallucination detection	🔗
Agentic CLEAR (ACL 2026) ⭐	IBM Research	Multi-level agent eval	Yes (OSS)	Free	Automated multi-level evaluation; dynamic issue discovery; MLflow/Langfuse integration	🔗
Styxx ⭐	Cognitive observability	Yes (MIT)	Free	9 cognometric instruments; hallucination/refusal/tool-call drift/goal-drift detection; per-step localization	🔗
Aether ⭐	Local-first cognition debugger	Yes	Free	Chrome DevTools for AI; real-time reasoning tree inspection; VS Code extension; hallucination debugging	🔗
Dynatrace dt-evals ⭐	Dynatrace	LLM/agent eval from traces	Yes (OSS)	Free (OSS); cloud paid	Eval from real GenAI traces; LLM judge; CI/CD integration; supports OpenAI/Anthropic/Google/AWS/Azure	🔗
Currai ⭐	Currai	AI observability platform	No	Paid	Prompt tracing; A/B testing; LLM evaluations; cost analytics; OpenTelemetry	—
tracesage ⭐	Open-source	LangGraph observability	Yes (MIT)	Free	Local-first; interactive graph + timeline UI; MCP tool-source attribution; pytest fixture	🔗
Opik ⭐	Comet	LLM lifecycle platform	Yes (Apache 2.0)	Free (OSS); cloud paid	Evaluation, testing, monitoring, optimization; Opik Guardrails	🔗
Orq.ai ⭐	Orq.ai	Observability + monitoring	No	Paid	Real-time monitoring; automated evaluations; trace automation; custom dashboards	—

MCP Ecosystem 🔌

The Model Context Protocol (MCP) is an open standard originally by Anthropic, now governed by the Agentic AI Foundation (AAIF) under the Linux Foundation (co-founded by Anthropic, Block, and OpenAI; supported by Google, Microsoft, AWS, Cloudflare, Bloomberg). It connects LLMs to external tools and data sources via a unified JSON-RPC 2.0 interface, supporting STDIO and Streamable HTTP transports. The official MCP Registry at registry.modelcontextprotocol.io is currently in preview as the centralized metadata repository for publicly accessible MCP servers, with registry docs at modelcontextprotocol.io/registry. Official registry docs now explicitly position it as metadata infrastructure for downstream aggregators and compatible registries, not the primary direct integration point for host apps. The 2026-07-28 spec release candidate (locked May 21, 2026) introduces stateless core transport, Tasks extension for long-running work, MCP Apps for server-rendered UIs, OAuth 2.1 alignment, a formal deprecation policy, and deprecates legacy roots and sampling features on the new spec line. Docker Custom MCP Catalogs and Profiles reached GA in May 2026. Microsoft Power BI MCP server released June 2026.

MCP Clients: Claude Desktop, Claude Code, Cursor, Windsurf, VS Code (Copilot + ACP Client), Continue.dev, Zed, ChatGPT, Gemini, Microsoft Copilot, LibreChat, and more.

Popular MCP Servers

Tool / Server	Developer	Category	Open Source	Best For	GitHub
MCP Filesystem	Anthropic / Community	File I/O	Yes (MIT)	Read/write local files from any MCP client	🔗
MCP GitHub	GitHub / Anthropic	Code & DevOps	Yes	Repo management, issues, PRs, code search	🔗
MCP Slack	Community	Messaging	Yes	Slack workspace read/write interaction	🔗
MCP PostgreSQL	Community	Database	Yes	Read-only SQL queries against Postgres	🔗
MCP Google Drive	Community	Storage	Yes	Drive file access and search	🔗
MCP Docker	Community	DevOps	Yes	Container management and inspection	🔗
MCP Brave Search	Brave	Search	Yes	Web + local search via Brave API	🔗
MCP AWS	AWS Labs	Cloud	Yes (Apache 2.0)	AWS service integration	🔗
MCP Notion	Community	Productivity	Yes	Notion page and database access	🔗
FastMCP	Community	Framework	Yes	Python framework for building MCP servers fast	🔗
Context7	Upstash	Dev Tools	Yes	Up-to-date library docs for AI coding assistants	🔗

MCP 2026-07-28 Release Candidate (May 21, 2026): The largest revision since launch. Key changes: stateless core (no sessions, no initialize handshake, plain HTTP load balancing), MCP Apps extension (server-rendered UIs), Tasks extension (long-running work), OAuth 2.0/OpenID Connect alignment, 12-month deprecation policy. Final spec ships July 28, 2026. This is a breaking change -- servers using session state must migrate to explicit handles. Source: 🔗 MCP SDK beta support (2026-06-29): Python, TypeScript, Go, and C# SDK beta releases now support the 2026-07-28 release candidate, so server maintainers can test stateless transport, Tasks, MCP Apps, and auth changes before final. Source: 🔗

Agent Skills & Registries 🎯

Modular capability packages that extend AI agents with specialized knowledge, workflows, and procedural instructions — without bloating model context.

skills.sh

skills.sh is the primary registry and package manager for Agent Skills — an open standard developed by Anthropic for packaging and distributing reusable agent capabilities. Skills follow a progressive disclosure pattern: agents load only a skill's name and description at startup, then pull full instructions only when a task matches, keeping context overhead minimal.

Vercel's official Agent Resources now publish installable skills for React, Next.js, AI SDK, design/UI, browser automation, deployment, commerce, and workflow use cases. Source: 🔗

Feature	Detail
Standard	Agent Skills (open, SKILL.md format) — developed by Anthropic, hosted on GitHub
Registry URL	skills.sh
Total installs	90,989+ all-time
Compatible agents	Claude Code, Cursor, Windsurf, VS Code Copilot, Continue.dev, Zed, and any MCP-compatible agent
License	Open (skills are author-licensed; spec is open standard)

Top Skills by Category

Skill	Publisher	Category	Installs
find-skills	vercel-labs/skills	Discovery	1.3M
vercel-react-best-practices	vercel-labs/agent-skills	Frontend	366K
frontend-design	anthropics/skills	Design	361K
web-design-guidelines	vercel-labs/agent-skills	Design	291K
microsoft-foundry	microsoft/azure-skills	Cloud/Azure	286K
azure-ai	microsoft/azure-skills	AI/Cloud	276K
agent-browser	vercel-labs/agent-browser	Browser	229K
skill-creator	anthropics/skills	Meta	180K
browser-use	browser-use/browser-use	Automation	71.6K
systematic-debugging	obra/superpowers	Dev	78.5K
test-driven-development	obra/superpowers	Dev	68.0K
seo-audit	coreyhaines31/marketingskills	Marketing	95.4K
supabase-postgres-best-practices	supabase/agent-skills	Database	138K
playwright-best-practices	currents-dev/playwright	Testing	34.2K

Notable Publisher Ecosystems

Publisher	Skills Count	Focus
microsoft/azure-skills	19+	Azure cloud, AI, Kubernetes, cost optimization
vercel-labs/agent-skills	15+	React, Next.js, Tailwind, deployment
anthropics/skills	15+	Design, docs, coding, web artifacts
coreyhaines31/marketingskills	20+	SEO, marketing, content, analytics
obra/superpowers	12+	Dev workflows, parallel agents, TDD
firebase/agent-skills	10+	Firebase, Firestore, GenKit
larksuite/cli	13+	Lark workspace automation
pbakaus/impeccable	10+	Design polish, code quality

Model Routers & Load Balancers 🔀

Tools for routing LLM requests across multiple providers, models, and deployments — optimizing for cost, latency, quality, or reliability. Prices as of July 2026.

Tool	Developer	Key Features	Open Source	Pricing	GitHub
LiteLLM	BerriAI	100+ provider support, proxy server, load balancing, fallbacks, spend tracking	Yes (MIT)	Free (OSS) / $99/mo cloud	🔗
Portkey	Portkey	250+ LLMs, AI gateway, guardrails, observability, virtual keys	Yes (Apache 2.0)	Free tier / $49/mo+	🔗
OpenRouter	OpenRouter	200+ model catalog, unified API, pay-per-use credit system	No	~5% markup on provider cost	—
RouteLLM	LMSys	Open-source router (strong vs. weak model) using classifier or matrix factorization	Yes	Free	🔗
Not Diamond	Not Diamond	Pre-trained + custom task-specific routers, cost/quality tradeoff	No	Free tier + enterprise	—
Unify AI	Unify	Quality / cost / latency-aware routing across 100+ model deployments	No	Usage-based	—
Semantic Router	Aurelio AI	Embedding-based semantic intent routing for agents and pipelines	Yes	Free	🔗

Small Language Models (SLMs) 📱

Compact models designed for on-device inference, edge deployment, low-latency APIs, and resource-constrained environments. Generally defined as models under ~15B parameters. Specs as of July 2026.

Model	Developer	Params	Context	License	Best For
Phi-4	Microsoft	14B	16K	MIT	Reasoning, math, code — STEM benchmark leader at class size
Phi-4-mini	Microsoft	3.8B	128K	MIT	On-device STEM reasoning with long context
Phi-4-multimodal	Microsoft	5.6B	128K	MIT	Vision + audio + text multimodal, edge deployment
Gemma 4 31B	Google	31B dense	256K	Apache 2.0	Multimodal (text+image), hybrid-thinking, open-weight reasoning
Gemma 4 26B A4B	Google	26B MoE (4B active)	256K	Apache 2.0	Multimodal, efficient MoE, agentic workflows
Gemma 4 E4B	Google	4B dense	256K	Apache 2.0	On-device, CPU inference, multimodal (text+image+audio)
Gemma 4 E2B	Google	2B dense	256K	Apache 2.0	Ultra-lightweight, on-device, 5GB RAM (4-bit)
Gemma 4 12B	Google	12B dense	256K	Apache 2.0	Encoder-free unified multimodal (text+image+audio), hybrid-thinking
Aion 1.0 Instruct ⭐	Microsoft	—	—	Open weights (July 2026)	On-device SLM for Windows; smaller/faster than current Windows SLM; Edge Insider preview; open-source on HuggingFace July 2026
Aion 1.0 Plan ⭐	Microsoft	14B	32K	—	On-device reasoning/tool-calling SLM; agentic workflows locally; shipping in-box Windows "in the coming months"
edge-lm (Gemma 4 E2B/E4B) ⭐	TheStageAI	2B/4B	256K	MIT	7x smaller Gemma 4 checkpoints; E2B in 1.44GB; MLX-ready; optimized for Apple Silicon/edge
Atome LM ⭐	TilelliLab	60K	—	Apache 2.0	Ternary zero-heap LM for microcontrollers; runs on ESP32; ~1 tok/s on Cortex-M3
Gemma 3 27B	Google	27B	128K	Apache 2.0	Top open model, multilingual (140+ languages)
Gemma 3 4B	Google	4B	128K	Apache 2.0	CPU inference, 140+ languages, mobile-friendly
Gemma 3 1B	Google	1B	32K	Apache 2.0	On-device, embedded, ultra-lightweight
Qwen3.5-9B	Alibaba	9B	256K	Apache 2.0	Thinking + non-thinking modes; native multimodal; 201 languages; reasoning closes gap to 30B+ models
Qwen3.5-4B	Alibaba	4B	256K	Apache 2.0	Multimodal (native VLM), agentic, on-device; edge deployment
Qwen3.5-2B	Alibaba	2B	256K	Apache 2.0	Runs on recent iPhones (offline); 201 languages; text+image
Qwen3.5-0.8B	Alibaba	0.8B	256K	Apache 2.0	IoT / rapid prototyping; non-thinking mode only; ultra-lightweight
Mistral Small 4	Mistral AI	119B total / 6B active (MoE, 128 experts)	256K	Apache 2.0	Chat + reasoning + coding + vision (multimodal) in one model; released 2026-03-16
SmolLM3	Hugging Face	3B	128K	Apache 2.0	Efficient, tool use, multilingual, reasoning
Qwen2.5 3B	Alibaba	3B	128K	Apache 2.0	Asian and multilingual tasks, coding
Qwen2.5 7B	Alibaba	7B	128K	Apache 2.0	Strong multilingual baseline, function calling
Llama 3.2 3B	Meta	3B	128K	Llama 3.2 license	General-purpose, on-device, Meta ecosystem
Llama 3.2 1B	Meta	1B	128K	Llama 3.2 license	Lightweight edge inference, distillation target
Granite 3.3 8B	IBM	8B	128K	Apache 2.0	Enterprise tasks, tool use, business-domain
MiniCPM 3.0	ModelBest / Tsinghua	4B	32K	Apache 2.0	Compact yet capable, mobile and edge
Danube 3 500M	H2O.ai	500M	8K	Apache 2.0	Ultra-lightweight on-device, IoT
MobileMoE ⭐	Academic	0.3B–0.9B active (1.3B–5.3B total)	128K	Research	First sub-billion MoE family for on-device; 1.8-3.8x faster than dense baselines
SLM-10M (Liodon AI) ⭐	Liodon AI	9.97M	1K	Apache 2.0	Leads sub-10M Open SLM Leaderboard (32.38%); beats Pythia-31M with 1/3 params
MiniGPT-16MB ⭐	david-spies	1.46M	256	MIT	Sub-16MB model; runs in browser via WebAssembly; 60-120+ tok/s
Micro Language Models (muLMs) ⭐	Sensente	8M–30M	—	Research	Instant on-device response initiation; masks cloud latency; collaborative generation
PhoneLM ⭐	UbiquitousLearning	0.5B / 1.5B	32K / 128K	Apache 2.0	Smartphone-native SLM; architecture searched for NPU efficiency; Android intent invocation

Apple Foundation Models 3rd Generation (June 8, 2026): Apple announced five foundation models powering Apple Intelligence. AFM 3 Core (3B dense) and AFM 3 Core Advanced (20B sparse, 1-4B active) run on-device. AFM 3 Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro run on Private Cloud Compute. Not available via public API. AFM 3 Core Advanced achieves 4.15 MOS for TTS (vs 3.87 production baseline) and 44.7% preference on dictation quality. Source: 🔗

Notable GitHub repos:

SmolLM / SmolLM2: 🔗
Granite 3.x: 🔗
MiniCPM: 🔗
Danube 3: 🔗
MobileMoE: 🔗
SLM-10M: 🔗

Guides 📚

Tutorials, how-tos, and in-depth guides for getting the most out of AI models and tools.

Getting Started 🚀

A beginner-friendly introduction to AI models and how to start using them effectively.

Understanding LLMs

Concept	Description
Parameters	Size of model (B = billions). More = more capable
Context Window	How much text model can process (128K standard)
Tokens	Basic units of text (~0.75 words per token)

Accessing AI Models

Method	Best For	Setup Difficulty
Web Interfaces	Quick experiments	Easiest
API Access	Building applications	Easy
Self-Hosting	Privacy, no API costs	Medium-Hard
IDE Integration	Daily coding	Easy

Model Recommendations by Task

Task	Free Option	Premium Option
Chat	Llama 4 (self-hosted)	Claude Fable 5, GPT-5.5 Instant
Coding	GLM-5.2, Qwen3-Coder (self-hosted)	Claude Fable 5, Claude Opus 4.8
Reasoning	DeepSeek-R1	Gemini 3 Deep Think, GPT-5.5 Pro
Long docs	Llama 4 Scout	Gemini 3 Flash
Vision	Llama 4 Maverick	GPT-5.5, Gemini 3.1 Pro

Free Models & APIs for Vibe Coding 💻

Vibe coding — describing what you want in natural language and letting AI generate the code — has exploded in 2026. The ecosystem splits into two tracks: free AI APIs you plug into your own editor/agent, and free vibe coding IDEs/platforms that bundle everything together.

Free AI APIs for Coding

These are the raw API endpoints you can use in tools like Cursor (BYOK), Cline, or any agent framework.

Provider	Free access	Paid baseline if you outgrow free	Best For	Source
Google Gemini API	Gemini Developer API free tier is free of charge	Gemini 2.5 Flash-Lite is $0.10 / $0.40 standard; $0.05 / $0.20 in Batch/Flex	Long-context and multimodal coding helpers	🔗
Groq Cloud	GroqCloud Free plan	Llama 3.1 8B Instant is $0.05 / $0.08 on paid usage	Very fast code loops and agent backends	🔗, 🔗
Cohere	Trial keys are free but limited; Command A+ is free until rate limits are reached	Command R is $0.15 / $0.60	Tool use, RAG, and multilingual coding support	🔗, 🔗, 🔗
Mistral AI	Free mode is enabled by default with no credit card required	Mistral Small 4 is $0.15 / $0.60 on Scale	Cheap coding and multimodal experiments	🔗, 🔗
Anthropic	New users receive a small amount of free credits	Haiku 4.5 is $1.00 / $5.00	Claude-first agent and prompt testing	🔗, 🔗
Vercel AI Gateway ⭐	$5/month included on the free tier for eligible models	Paid usage follows provider list rates with zero markup	Multi-provider prototyping behind one endpoint	🔗
OpenAI	Developer quickstart includes one free test API request	GPT-5.4 nano is $0.20 / $1.25	OpenAI SDK integration checks	🔗, 🔗

Free Vibe Coding IDEs & Platforms

Tool	Type	Key Features	Best For
Cursor	AI IDE	Agent Mode, Composer 2, multi-agent workspace	Professional development
Cline	VS Code Extension	Open-source, BYOK/Ollama, MCP tools	Self-hosted, unlimited local LLM
Windsurf	AI IDE	Cascade agent, live browser preview	IDE with browser integration
OpenHands	Docker Agent	Self-hosted, local LLM support, full SDLC	Unlimited local development
bolt.diy	Browser IDE	19+ LLM providers, Ollama, full-stack apps	Free web app building
Open Interpreter	CLI	Natural language → code, local LLM	Simple local automation

Chrome DevTools MCP - Game Changer for Web Dev

Google's Chrome DevTools MCP connects AI agents directly to Chrome for debugging, profiling, and automation:

29 tools across 6 categories (input, navigation, emulation, performance, network, debugging)
Run Lighthouse audits, capture performance traces, inspect network requests
Works with Claude Code, Cursor, Copilot via MCP
Supports BYOK/local LLMs through MCP clients
GitHub | 37,783+ stars

Cloudflare Browser Run

Managed browser infrastructure for AI agents:

Chrome DevTools Protocol (CDP) direct endpoint
MCP client support (Claude, Cursor, OpenCode)
Session recordings, Live View, WebMCP, and /json schema extraction with Workers AI or BYOK
Free Workers plan / $5+/mo paid
Browser Run

Recommendations by Use Case

Free + Local LLM: Cline + Ollama, OpenHands + Qwen3 Coder, bolt.diy + Ollama

Fast API Iteration: Groq (speed) + Cerebras (high limits)

Web Development: Chrome DevTools MCP + Cline (zero-cost debugging)

Many Models: OpenRouter (unified API)

Production Inference: NVIDIA NIM, Cerebras

💡 Pro Tip: Combine Chrome DevTools MCP with a local LLM (Ollama) via Cline for completely free, unlimited AI-powered web development and debugging.

A comprehensive guide to running AI models on your own hardware.

Benefits

Benefit	Description
Privacy	Data never leaves your infrastructure
Cost Control	No per-token API costs for unlimited usage
Customization	Fine-tune models for specific needs
No Rate Limits	Process as much as hardware allows
Offline Access	Work without internet

Quick Start with Ollama

For installation and usage instructions, refer to the official Ollama documentation.

Local GPU Quick Guide

Recommended apps (local-first):

Ollama - Simple local runtime with a local HTTP API
LM Studio - Desktop UI for downloading and running models locally
llama.cpp - Fast local inference (CPU/GPU), great for quantized models
Open WebUI - Optional local web UI (pairs well with local runtimes)

If you want “server-style” hosting (advanced):

vLLM - High-throughput serving for NVIDIA GPUs
SGLang - Structured generation and serving workflows

Practical setup tips:

Install the latest NVIDIA drivers (enable GPU acceleration in your chosen app)
Start with smaller quantized models (Q4 is a common “best default”)
Keep context windows realistic for local hardware (lower context = faster, less memory)
Watch VRAM first, then system RAM; reduce model size or quantization if either saturates
Prefer running locally on localhost and only expose to LAN if you understand firewall rules

Example hardware configurations:

Hardware	Good starting point	Notes
Consumer GPU (24 GB VRAM)	7B–14B quantized	e.g., RTX 4090, RTX 3090 — great for chat/coding
Pro GPU (48–80 GB VRAM)	14B–70B quantized	e.g., A6000, A100 — coding agents, longer contexts
Multi-GPU (160+ GB VRAM)	70B+ quantized	e.g., 2×A100 — larger open-source models
CPU-only (32–64 GB RAM)	7B–14B quantized	Slower but viable for offline chat; keep context moderate

Deployment Options

Option	Best For	Pros	Cons
Local Machine	Personal use	Simple, no latency	Limited hardware
Dedicated Server	Team use	Full control	Maintenance
Cloud GPU Rental	Experimentation	On-demand	Hourly costs
Kubernetes	Enterprise	Scalable	Complex

Cost Analysis 💰

Comprehensive pricing comparisons and cost calculations.

Pricing Tiers

Tier	Price Range	Models
🆓 Free	$0	Self-hosted, free tiers
💸 Budget	$0.025 - $0.50/1M	Gemini 3.1 Flash-Lite, GPT-4.1 nano, GLM-4.7-FlashX, GPT-5.4 nano, Grok 4 Fast (aliased to 4.3)
💰 Mid-range	$0.60 - $15.00/1M	GPT-5.4 mini, Claude Haiku 4.5, Kimi K2.6, Sonar, GLM-5, GPT-5.4, Claude Sonnet, Grok 4.3
💎 Premium	$15.00 - $600.00/1M	Claude Fable 5, GPT-5.5 Pro, Claude Opus 4.8, Claude Opus 4.7, o1-Pro

Subscription Pricing (Monthly, USD)

AI chat apps

Product	Plans (USD)	Notes	Official Source
ChatGPT	Go $8, Plus $20, Pro $200, Business $25/seat (annual) or $30/seat (monthly), Enterprise (contact sales)	Consumer prices are US-listed; Go is localized in some markets	🔗
Claude	Pro $20, Max $100 (5×) or $200 (20×), Team/Enterprise (see pricing)	Prices shown exclude applicable taxes; availability varies by region	🔗
Google AI (Gemini)	Free, Plus $4.99, Pro $19.99, Ultra $99.99 (5× limits / 20TB storage) or $200 (20× limits / 20TB storage)	US pricing; Plus price dropped from $7.99 on June 8, 2026 with storage doubled to 400GB; Ultra reduced from $249.99 at I/O 2026; $99.99 tier added with Gemini 3.5 Flash integration and Antigravity priority access	🔗

Coding assistants

Tool	Plans (USD)	Notes	Official Source
GitHub Copilot	Free $0, Pro $10 base + $5 monthly bundled AI Credits, Pro+ $39 base + $31 bundled AI Credits, Business $19/user, Enterprise $39/user	Usage-based billing since June 1, 2026; chat, agent mode, code review, cloud agent, CLI, and apps draw from GitHub AI Credits	🔗

Model Pricing Comparison

Comparative pricing across the cheapest official public APIs verified in this pass. Only providers and models rechecked against current official pricing pages are included below. As of 2026-07-19 16:54 UTC.

Model	Input	Output	Cached Input	Best For	Source
Llama 3.1 8B Instant (Groq)	$0.05	$0.08	—	Lowest verified public direct text API price	🔗
GPT OSS 20B (Groq)	$0.075	$0.30	$0.0375	Cheap open-weight routing and agents	🔗
Gemini 2.5 Flash-Lite	$0.10	$0.40	$0.01	Cheapest verified multimodal API with an always-on free tier	🔗
DeepSeek-V4-Flash	$0.14	$0.28	$0.0028 (hit)	Strong price/performance for long-context work	🔗
Mistral Small 4	$0.15	$0.60	—	Low-cost coding and multimodal work	🔗
Command R (Cohere)	$0.15	$0.60	—	RAG and tool use	🔗
GPT-5.4 nano	$0.20	$1.25	$0.02	OpenAI lowest-cost standard model	🔗
Claude Haiku 4.5	$1.00	$5.00	$0.10 (read)	Lowest-cost Claude model	🔗
GPT-5.6 Luna	$1.00	$6.00	$0.10	Cheapest current GPT-5.6 tier	🔗
Kimi K3 ⭐	$3.00	$15.00	$0.30 (hit, 90% cached)	New open-weight frontier model (2026-07-16); 2.8T MoE, 1M context, native vision	🔗
Meta Muse Spark 1.1 ⭐	API-only	API-only	—	Closed-weights frontier multimodal reasoning; agentic tool and computer use; 1M context; released 2026-07-09	🔗
Grok-4.5	$2.00	$6.00	$0.50 (hit)	xAI new flagship (2026-07-08); 500K context, text+image; function calling, structured outputs, reasoning	🔗
Grok 4.3	$1.25	$2.50	$0.20	xAI flagship at the current public rate card	🔗
Claude Sonnet 5	$2.00 intro / $3.00 standard	$10.00 intro / $15.00 standard	$0.20 intro read	Strong coding model if the budget can stretch	🔗

Self-Hosting vs API (Monthly)

Usage Level	Self-Host (A100)	API (GPT-5)	Winner
Light (1M tokens)	$300 (rental)	$10	API
Medium (100M tokens)	$300	$1,000	Self-host
Heavy (1B tokens)	$300	$10,000	Self-host
Enterprise (10B+ tokens)	$2,000 (owned)	$100,000+	Self-host

Reference 📖

Reference materials including glossary, comparison tables, and data sources.

Glossary 📖

Definitions of common terms used throughout the documentation.

A-E

Term	Definition
Agent	AI system that autonomously performs tasks and interacts with environments
API	Interface for programmatically accessing AI models
Attention Mechanism	Neural network component focusing on relevant input parts
Benchmark	Standardized test measuring model performance
Chain-of-Thought (CoT)	Prompting technique showing step-by-step reasoning

F-L

Term	Definition
Fine-Tuning	Adapting pre-trained model to specific tasks
Frontier Model	State-of-the-art proprietary model
GPU	Hardware accelerator essential for ML
LLM	Large Language Model
LoRA	Efficient fine-tuning method

M-R

Term	Definition
MCP	Model Context Protocol for tool interaction
MMLU	Massive Multitask Language Understanding benchmark
MoE	Mixture of Experts architecture
Multimodal	Processing multiple input types
RAG	Retrieval-Augmented Generation

S-Z

Term	Definition
Self-Hosting	Running models on own infrastructure
SLM	Small Language Model
SWE-bench	Benchmark for real GitHub issue resolution
Token	Basic unit of text processing
VRAM	GPU memory for model storage

Comparison Tables 📊

Side-by-side comparisons of AI models sorted by various criteria.

Sort by Latest Update (Default)

🏢 Company	🤖 Model	📦 Version	📅 Release Date	🔄 Latest Updated	💻 Coding	📊 Benchmarks	💰 Price	🖥️ Self-Host	🔗 Official Site
🤖 Anthropic	Claude	Fable 5 redeployed	2026-06-09 00:00 UTC	2026-07-01 00:00 UTC ⭐	✅	SWE-bench 95.0%	$10.00 / $50.00	❌	🔗
🤖 Anthropic	Claude	Sonnet 5	2026-06-30 00:00 UTC	2026-06-30 00:00 UTC ⭐	✅	SWE-bench 92.4%, SWE-bench Pro 63.2%, HLE 57.4% with tools	$2.00 / $10.00 intro	❌	🔗
🤖 OpenAI	GPT-5.6	Sol / Terra / Luna	2026-07-09 00:00 UTC	2026-07-09 00:00 UTC ⭐	✅	General availability; Sol $5/$30, Terra $2.50/$15, Luna $1/$6	❌	🔗
🇨🇳 Moonshot AI	Kimi	K3	2026-07-16 00:00 UTC	2026-07-16 00:00 UTC ⭐	✅	2.8T MoE open-weight frontier; 1M context; $3.00/$15.00; 90% cache-hit discount	✅	🔗
🇨🇳 Meta	Meta	Muse Spark 1.1	2026-07-09 00:00 UTC	2026-07-09 00:00 UTC ⭐	✅	Closed-weights frontier multimodal reasoning; agentic tool and computer use; 1M context; API-only	❌	🔗
🚀 xAI	Grok	4.5	2026-07-08 00:00 UTC	2026-07-08 00:00 UTC ⭐	✅	New frontier multimodal LLM; 500K context; $2.00/$6.00	❌	🔗
🤖 OpenAI	GPT-Live	1 / 1 mini	2026-07-08 00:00 UTC	2026-07-08 00:00 UTC ⭐	—	Full-duplex voice models powering ChatGPT Voice; GPT-5.5 backend; API pending	❌	🔗
🇨🇳 MiniMax	M2 / M2.1	Open-source	2026-07-02 00:00 UTC	2026-07-02 00:00 UTC ⭐	✅	230B/10B active MoE, 1M context, Apache 2.0	✅	🔗
🌐 Google DeepMind	Gemini Omni Flash	—	2026-06-30 00:00 UTC	2026-06-30 00:00 UTC ⭐	✅	$0.10/s video; any-to-any multimodal	❌	🔗
🌐 Google DeepMind	Nano Banana 2 Lite	—	2026-06-30 00:00 UTC	2026-06-30 00:00 UTC ⭐	—	$0.034/1K images; 4s generation	❌	🔗
🇨🇳 Zhipu AI	GLM-5.2	—	2026-06-13 00:00 UTC	2026-06-17 00:00 UTC ⭐	✅	AIME 2026 99.2%, SWE-bench Pro 62.1%, FrontierSWE 74.4%	$1.40 / $4.40	✅	🔗
🌐 Google DeepMind	Gemini 3.5	Flash	2026-05-19 00:00 UTC	2026-05-19 00:00 UTC ⭐	✅	GPQA ~90.4%, SWE-bench Pro 55.1%	$1.50 / $9.00	❌	🔗
🤖 Anthropic	Claude	Mythos Preview	2026-04-07 00:00 UTC	2026-04-07 00:00 UTC	—	Not disclosed	$25.00 / $125.00	❌	🔗
🔬 DeepSeek	DeepSeek	V4 (Flash/Pro)	2026-04-24 00:00 UTC	2026-05-08 00:00 UTC	✅	No public benchmarks	From $0.14 / $0.28 (Flash)	✅	🔗
🤖 OpenAI	GPT-5.5	Pro	2026-04-26 00:00 UTC	2026-05-08 00:00 UTC	✅	GPQA 95.1%, SWE-bench 92.3%	$30.00 / $180.00	❌	🔗
🌐 Google DeepMind	Gemini 3	Flash	2026-02-12 00:00 UTC	2026-05-08 00:00 UTC	✅	GPQA 90.4%, SWE-bench 78.0%	$0.50 / $3.00	❌	🔗
🚀 xAI	Grok	4.3	2026-05-01 00:00 UTC	2026-05-06 00:00 UTC	✅	—	$1.25 / $2.50	❌	🔗
🇨🇳 MiniMax	Hailuo	02	2026-06-20 00:00 UTC	2026-06-20 00:00 UTC ⭐	—	#2 Video Arena	API	❌	🔗
🇨🇳 MiniMax	Speech	2.8	2026-01-23 00:00 UTC	2026-07-03 21:56 UTC ⭐	—	Native sound tags, high-fidelity cloning, studio-grade clarity	API	❌	🔗
🇨🇳 MiniMax	Video-01	—	2026-06-20 00:00 UTC	2026-06-20 00:00 UTC ⭐	—	720p/25fps	~$0.40/video	❌	🔗
🇫🇷 Mistral AI	Voxtral TTS	—	2026-06-18 00:00 UTC	2026-06-18 00:00 UTC ⭐	—	4B, 9 languages, 70ms	$0.016/1K chars	✅	🔗
🤖 Microsoft	VibeVoice Realtime (0.5B) ⭐	—	2025-12-03 00:00 UTC	2026-03-06 00:00 UTC ⭐	—	0.5B real-time TTS; <300ms TTFA; streaming input	Free (MIT)	✅	🔗
🤖 Microsoft	VibeVoice ASR (7B) ⭐	—	2026-01-21 00:00 UTC	2026-03-06 00:00 UTC ⭐	—	60-min ASR with diarization + timestamps	Free (MIT)	✅	🔗
🤖 OpenAI	GPT-4.1 ⭐	—	2026-04-14 00:00 UTC	2026-04-14 00:00 UTC	✅	1M context, improved coding/instruction following	$2.00 / $8.00	❌	🔗
🤖 OpenAI	GPT-4.1 mini ⭐	—	2026-04-14 00:00 UTC	2026-04-14 00:00 UTC	✅	1M context, best value for large-context needs	$0.40 / $1.60	❌	🔗
🇺🇸 Zyphra	ZONOS2	—	2026-06-12 00:00 UTC	2026-06-12 00:00 UTC ⭐	—	8B MoE TTS, multilingual	Free (Apache 2.0)	✅	🔗
🤖 OpenAI	GPT-5.5	—	2026-04-26 00:00 UTC	2026-04-26 00:00 UTC	✅	GPQA 93.2%, SWE-bench 88.5%	$5.00 / $30.00	❌	🔗
🤖 Anthropic	Claude	Opus 4.7	2026-04-22 00:00 UTC	2026-04-26 00:00 UTC	✅	GPQA 94.2%, SWE-bench 87.6%	$5.00 / $25.00	❌	🔗

Release Windows (Month-level)

🏢 Company	🤖 Model	📅 Release Window	Notes	🔗 Official Site
🤖 OpenAI	GPT-5.6	2026-07	Sol/Terra/Luna three-tier family; general availability; Sol $5/$30, Terra $2.50/$15, Luna $1/$6	🔗
🇨🇳 Moonshot AI	Kimi K3	2026-07	2.8T MoE open-weight frontier; 1M context; $3.00/$15.00; 90% cache-hit discount; open weights promised by 2026-07-27	🔗
🇨🇳 Meta	Meta Muse Spark 1.1	2026-07	Closed-weights frontier multimodal reasoning; agentic tool and computer use; 1M context; API-only (Meta Model API)	🔗
🇨🇳 MiniMax	M2 / M2.1	2026-07	Open-source MoE (230B/10B active); 1M context; Apache 2.0; $0.30/$1.20 API	🔗
🌐 Google DeepMind	Gemini Omni Flash	2026-06	$0.10/s video generation; any-to-any multimodal; conversational editing	🔗
🌐 Google DeepMind	Nano Banana 2 Lite	2026-06	Fastest Gemini Image; 4s gen; $0.034/1K images	🔗
🤖 Anthropic	Claude	Fable 5	2026-06	First publicly available Mythos-class model; $10/$50; free on subscription through June 22
🤖 Anthropic	Claude	Mythos 5	2026-06	Restricted to Project Glasswing partners; same model as Fable 5 with safeguards lifted
🇨🇳 Alibaba	Qwen3.7-Max	2026-05	Proprietary; $2.50 / $7.50	🔗
🇨🇳 Zhipu AI	GLM-5.2	2026-06	753B params, 1M context, MIT open-weight; AIME 2026 99.2%; API $1.40/$4.40	🔗
🌐 Google DeepMind	Gemini 3.5 Flash	2026-05	GA at I/O 2026; 4x faster than 3.1 Pro
🌐 Google DeepMind	Gemini Omni Flash	2026-05	Any-to-any multimodal creation; rolling out to AI Plus/Pro/Ultra	🔗
🤖 Anthropic	Claude Opus 4.8	2026-05	Improved Opus 4.7; SWE-bench 88.6%; fast mode $10/$50	🔗
🚀 xAI	Grok 4.20 / 4.3 / 4 Fast	2026-05	Major price cuts; Grok 4.20 now $1.25/$2.50; 4 Fast → 4.3 alias	🔗
🤖 Anthropic	Claude Mythos Preview	2026-04	Invitation-only via Project Glasswing	🔗
🤖 OpenAI	GPT-5.5 Instant	2026-05	Default ChatGPT model since 2026-05-05	🔗
🚀 xAI	Grok 4.3	2026-05	Always-on reasoning; $1.25 / $2.50	🔗
🧠 MiniMax	MiniMax M2.5	2026-02	$0.30 / $1.20	🔗
🇨🇳 Alibaba/Qwen	Qwen 3.5-Max	2026-02	Open-source release window	🔗
🌐 Google DeepMind	Gemini 3.1 Flash-Lite	2026-02	Budget Gemini model	🔗
🌐 Google DeepMind	Gemini 3 Pro	2026-01	Tiered pricing	🔗
🤖 OpenAI	GPT-5.4 family	2026-03	GPT-5.4, GPT-5.4 mini, GPT-5.4 nano	🔗
🤖 OpenAI	GPT-4.1 family ⭐	2026-04	GPT-4.1 ($2/$8), GPT-4.1 mini ($0.40/$1.60), GPT-4.1 nano ($0.10/$0.40); 1M context; June 2024 cutoff; retired from ChatGPT Feb 13, 2026; API still available	🔗
🇫🇷 Mistral AI	Mistral Large 3	2025-11	Apache 2.0 open-source, 123B params	🔗

Sort by Price (Cheapest)

Rank	Model	Input	Output	License
1	Self-hosted open weights	$0	$0	Various
2	Llama 3.1 8B Instant (Groq)	$0.05	$0.08	API
3	GPT OSS 20B (Groq)	$0.075	$0.30	Open weights via API
4	Gemini 2.5 Flash-Lite	$0.10	$0.40	Proprietary
5	DeepSeek-V4-Flash	$0.14	$0.28	API
6	Mistral Small 4	$0.15	$0.60	Open weights/API
7	Command R	$0.15	$0.60	API
8	GPT-5.4 nano	$0.20	$1.25	Proprietary
9	Kimi K3 ⭐	$3.00	$15.00	Open-weight (Modified MIT)
10	Claude Haiku 4.5	$1.00	$5.00	Proprietary
10	GPT-5.6 Luna	$1.00	$6.00	Proprietary

Sort by Performance (Coding)

Rank	Model	SWE-bench Verified	Self-Host
1	Claude Mythos 5	95.5%	❌
2	Claude Fable 5	95.0%	❌
3	Claude Sonnet 5 ⭐	92.4%	❌
4	GPT-5.5 Pro	92.3%	❌
5	Claude Opus 4.8	88.6%	❌
6	Claude Opus 4.7	87.6%	❌
7	GPT-5.3-Codex	85.0%	❌
8	GLM-5.2 ⭐	62.1% (SWE-bench Pro)	✅
9	Kimi K3 ⭐	88.3% (Terminal-Bench 2.1)	✅
10	Claude Opus 4.6	80.8%	❌
10	Gemini 3.1 Pro	80.6%	❌

Sort by Context Window

Rank	Model	Context	Best For
1	Gemini 3 Flash	10M	Entire libraries
2	Llama 4 Scout	10M	Long-document RAG
3	Grok 4.20	2M	Large codebases with full context
4	Gemini 3 Pro	1M+	Research papers
5	Gemini 3.1 Pro	1M	Complex multi-document analysis
6	GLM-5.2 ⭐	1M	Frontier open-weight coding + long context
7	Kimi K3 ⭐	1M	2.8T MoE open-weight frontier; native vision; $3.00/$15.00
8	Gemini 3.5 Flash	1M	Fast multimodal reasoning
8	Claude Opus 4.7	1M	Agentic coding with full codebase
9	GPT-5.5 Pro	1.05M	Premium reasoning with long context
10	Grok 4.3	1M	Reasoning with full project context

Data Sources 📚

Attribution, verification sources, and methodology.

Primary Sources

Company	Source	URL
OpenAI	Official Documentation	openai.com
OpenAI	ChatGPT agent release notes	help.openai.com
OpenAI	Model release notes	help.openai.com
OpenAI	API pricing	platform.openai.com
OpenAI	API pricing (overview)	openai.com/api/pricing
OpenAI	GPT-5.5 pro model reference	developers.openai.com
OpenAI	GPT-5.3-Codex model reference	developers.openai.com
OpenAI	March 2026 model news	openai.com
OpenAI	ChatGPT subscriptions (Go/Plus/Pro)	openai.com
OpenAI	ChatGPT Business pricing	help.openai.com
OpenAI	GPT-4.5 API deprecation	developers.openai.com/api/docs/deprecations
OpenAI	GPT-4.1 / GPT-4o retirement from ChatGPT	openai.com
OpenAI	GPT-Image-2 / ChatGPT Images 2.0	openai.com
OpenAI	Codex changelog and CLI docs	developers.openai.com
OpenAI	Workspace agents in ChatGPT	openai.com
OpenAI	GPT-Live voice models announcement	openai.com
OpenAI	Realtime API (gpt-realtime 2.1) docs	platform.openai.com
OpenAI	Developer quickstart (free test request)	developers.openai.com
Anthropic	Claude Documentation	anthropic.com
Anthropic	Claude API pricing	docs.anthropic.com
Anthropic	Claude rate card and plan pricing	claude.com
Anthropic	Claude Sonnet 5 announcement	anthropic.com
Anthropic	Claude Sonnet 5 system card	anthropic.com
Anthropic	Claude Fable 5 redeployment	anthropic.com
Anthropic	Claude Opus 4.8 announcement	anthropic.com
Anthropic	Claude Fable 5 / Mythos 5 announcement	anthropic.com
Anthropic	Claude Tag announcement	anthropic.com
Anthropic	Claude Science workbench	anthropic.com
Anthropic	Claude Haiku 4.5 announcement	anthropic.com
Anthropic	Claude Pro pricing	anthropic.com
Anthropic	Max plan pricing	anthropic.com
Anthropic	Claude model selection and deprecations	docs.anthropic.com
Google	Gemini Documentation	deepmind.google
Google Cloud	Vertex AI Gemini pricing	cloud.google.com
Google	Gemini Developer API pricing	ai.google.dev
Google	Gemini API models (Flash-Lite pricing)	ai.google.dev
Google	Project Mariner	deepmind.google
Google	Google AI plans	one.google.com
Google	Google AI Plus pricing	blog.google
Google	Google AI Pro pricing	one.google.com
Google	Google AI Ultra pricing	blog.google
Google	Gemini API model catalog	ai.google.dev
Google	Gemini image model cards	deepmind.google
GitHub	Copilot plans & pricing	github.com
GitHub	Copilot changelog	github.blog
Kiro	IDE changelog	kiro.dev
Cursor	Pricing and changelog	cursor.com, cursor.com
Windsurf	Pricing and usage docs	windsurf.com, docs.windsurf.com
Zhipu AI (Z.ai)	Developer Documentation	docs.z.ai
Zhipu AI (Z.ai)	GLM-5.2 announcement	z.ai
DeepSeek	Models & API pricing; corporate site	api-docs.deepseek.com, deepseek.com
Mistral AI	Mistral Small 4 model card	docs.mistral.ai
Mistral AI	Free mode API key quickstart	docs.mistral.ai
MiniMax	Developer Documentation	platform.minimax.io
MiniMax	Pricing (Pay‑as‑you‑go)	platform.minimax.io
MiniMax	Speech 2.8 announcement	minimax.io
Moonshot AI	Developer Documentation	platform.moonshot.ai
Moonshot AI	Models & Pricing	platform.moonshot.ai
Moonshot AI	Kimi K3 announcement	kimi.com
Meta	Llama Documentation	llama.meta.com
Meta	Muse Spark 1.1 announcement	ai.meta.com
Cohere	Developer Documentation	docs.cohere.com
Cohere	Pricing overview and trial usage	docs.cohere.com
Cohere	Command R and Command A+ pricing	docs.cohere.com, docs.cohere.com
Groq	On-demand model pricing	groq.com
Groq	GroqCloud free and developer plans	groq.com
Cartesia	Sonic 3.5 / Ink-2 launch and docs	cartesia.ai, docs.cartesia.ai
AWS	Amazon Bedrock AgentCore	aws.amazon.com
Cloudflare	Browser Run and `/json` Quick Action	developers.cloudflare.com
Browser Use	Browser Use pricing, cloud docs, and releases	browser-use.com, docs.browser-use.com, github.com
Browser Use	Browser Harness	github.com
Vercel	agent-browser docs	agent-browser.dev
Vercel	AI SDK 7	vercel.com
Vercel	AI Gateway routing rules and pricing	vercel.com, vercel.com
Vercel	eve Agent Runs in MCP/CLI	vercel.com
Vercel	Sandbox FUSE support	vercel.com
Vercel	Agent Resources skills	vercel.com
Google	ADK 2.0 and ADK Go 2.0	adk.dev, developers.googleblog.com
Google	Genkit Agents	developers.googleblog.com
Mastra	File-based agents	mastra.ai
DeepReinforce	Ornith-1.0 model family	huggingface.co, huggingface.co
Jina AI	jina-embeddings-v5-omni	jina.ai
Model Context Protocol	Official registry and 2026-07-28 SDK betas	registry.modelcontextprotocol.io, blog.modelcontextprotocol.io
AI21 Labs	Developer Documentation	docs.ai21.com
Perplexity	Developer Documentation	docs.perplexity.ai
ByteDance (Volcengine)	Developer Documentation	volcengine.com
Tencent (Hunyuan)	Cloud Documentation	cloud.tencent.com
Baidu (ERNIE)	AI Studio Documentation	ai.baidu.com
xAI	Grok API pricing & models	docs.x.ai, docs.x.ai
Apple	Foundation Models 3rd Gen announcement	machinelearning.apple.com
Meta	Llama Documentation	llama.meta.com

Benchmark Sources

Benchmark	Source	Description
GPQA Diamond	Google Research	Graduate-level science questions (PhD difficulty)
MMLU-Pro	TIGER-Lab	Extended multi-task language understanding
Arena Elo	lmarena.ai	Crowdsourced human preference ranking
HLE	Scale AI	Humanity's Last Exam — expert-level questions
SWE-bench Verified	Princeton	Real GitHub issue resolution (human-verified)
SWE-bench Pro	Princeton	More challenging subset of SWE-bench
LiveCodeBench	LiveCodeBench	Live competitive programming problems
AIME 2025	MAA	American Invitational Mathematics Examination
ARC-AGI-2	ARC Prize	Abstract reasoning challenge (fluid intelligence)
MMMU / MMMU-Pro	MMMU	Multi-discipline multimodal understanding
IFEval	Google Research	Instruction-following evaluation
FrontierMath	Epoch AI	Expert-level research mathematics
HumanEval	OpenAI	164 Python programming problems

Verification Methodology

Primary Source Review - Check official documentation
Cross-Validation - Compare multiple sources
Timestamp Verification - All data includes verification date
Update Tracking - Monitor official channels

Supplemental agent & CLI notes 📎

Additional context from the June 2026 research pass (supplements earlier tables; nothing below removes or supersedes prior entries).

Google I/O 2026 agent announcements - Antigravity 2.0 (desktop + CLI + SDK) GA on 2026-05-19; Gemini Spark (24/7 personal agent) included with AI Ultra; Android CLI 1.0 stable for AI agent Android development; Gemini CLI sunset on 2026-06-18.
Microsoft Build 2026 - Windows Agent Framework open-sourced (MIT); Aion 1.0 Instruct and Aion 1.0 Plan on-device SLMs announced; Windows 365 for Agents GA; Surface RTX Spark Dev Box and DGX Station for Windows announced.
Claude Sonnet 5 launch - Anthropic launched Sonnet 5 on 2026-06-30 00:00 UTC with 1M context, 128K output, adaptive thinking, $2.00 / $10.00 introductory pricing through 2026-08-31 00:00 UTC, and 92.4% SWE-bench Verified.
Claude Fable 5 redeployment - Anthropic restored Fable 5 and Mythos 5 access on 2026-07-01 00:00 UTC after export controls were lifted; Fable 5 is available globally, while Mythos 5 remains limited to approved Project Glasswing partners.
Kimi K2.7 Code in GitHub Copilot - Moonshot AI's open-weight coding model became generally available in GitHub Copilot model picker on July 1, 2026; first open-weight model offered in Copilot; rolled out to Pro, Pro+, and Max plans.
Kimi K3 - Moonshot AI released Kimi K3 on 2026-07-16: 2.8T-parameter open-weight frontier model with 1M context, native vision, $3.00/$15.00 API pricing, 90% cache-hit discount ($0.30 cached input), and open weights promised by 2026-07-27 under a modified MIT license. Source: 🔗
Meta Muse Spark 1.1 - Meta released Muse Spark 1.1 on 2026-07-09: closed-weights frontier multimodal reasoning model built for agentic tool and computer use, 1M context, available via Meta Model API in public preview for US developers. Source: 🔗
MiniMax M2 & M2.1 open-source release - MiniMax open-sourced MiniMax M2 (230B/10B active MoE, 1M context, Apache 2.0) and M2.1 (229B/10B active MoE, 1M context, Apache 2.0) on July 2, 2026; API priced at $0.30/$1.20 per 1M tokens; available on Hugging Face, vLLM, SGLang, and MiniMax Agent.
Gemini Omni Flash and Nano Banana 2 Lite - Google released Gemini Omni Flash for video generation ($0.10/s) and Nano Banana 2 Lite for image generation (4s generation, fastest Gemini Image model) on June 30, 2026; available in Google AI Studio, Gemini API, and Gemini Enterprise Agent Platform.
GLM-5.2 open-source release - Zhipu AI open-sourced GLM-5.2 (753B MoE, 1M context, MIT license) on 2026-06-17; available via API at $1.40/$4.40 per 1M; leads AIME 2026 at 99.2%.
MCP 2026-07-28 spec RC - Largest revision since initial release; six breaking changes including stateless transport, two new required HTTP headers, Tasks lifecycle primitive; SDK maintainers have 10 weeks to ship support.
GPT-Live voice models - OpenAI launched GPT-Live (1, 1 mini, Medium, High) on 2026-07-08, replacing Advanced Voice Mode inside ChatGPT; full-duplex consumer voice stack with GPT-5.5 backend delegation. No developer API at launch (Realtime API with gpt-realtime-2.1, shipped 2026-07-06, is the API path). See the TTS Proprietary table entry.
Codex CLI 0.144.5 - OpenAI shipped a patch release on 2026-07-16 that expands dangerous-command detection, including more forced rm forms, and returns clearer denial reasons when a command is blocked.
Cursor 3.11 and Kiro 1.0.138 - Cursor added side chats, conversation search, and simpler project/repo pickers on 2026-07-10. Kiro's 2026-07-13 IDE release improved session startup, fixed compaction loops on large sessions, added full PowerShell trust on Windows, and improved MCP tool recovery after transient network failures.
skills.sh — Primary registry for Agent Skills packages (see Agent Skills & Registries 🎯); pairs naturally with CLI-first workflows (Claude Code, Gemini CLI, MCP-capable agents).
CrewAI — Multi-agent framework with OSS core plus AMP Studio / AMP Cloud paths; vendor positioning emphasizes transparent orchestration, visual workflows, broad tool integrations, and team-scale adoption (see the CrewAI row in the Multi-Agent / Parallel Agent Platforms table in this document).
Aiden — Listed above under local desktop agents as a local-first, Windows-native autonomous stack for running agents without mandatory cloud offload (taracodlabs/aiden).

Last Updated: 2026-07-19 16:54 UTC Maintained by: ReadyPixels LLC

📖 Additional Resources

Related Projects

Awesome AI - Curated list of AI resources
Awesome Machine Learning - ML frameworks and libraries
Hugging Face Open Source Models - Open-source model hub

Community

Discussions: GitHub Discussions
Issues: Report Issues

Support

📚 Documentation: Full Docs

📄 License

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

Made with ❤️ by ReadyPixels LLC

Name		Name	Last commit message	Last commit date
Latest commit History 174 Commits
.github		.github
assets		assets
docs		docs
.gitattributes		.gitattributes
.gitignore		.gitignore
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
SECURITY.md		SECURITY.md

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

Awesome AI Models Matrix 🧠

🚀 Quick Start

🔥 Highlights

📊 Top Models by Category

📚 Documentation

📋 At a Glance

🔍 Search & Navigation

Contents

Models 🧠

Frontier Models 🚀

Model Specifications 📋

Output Token Limits

Cached & Batch Pricing

Speed & Latency

Training Data Cutoffs

Multilingual Support

Structured Output & Function Calling

Regional Availability

Free-Source Models 🆓

Deployment Options

Coding Models 💻

SWE-bench Verified Leaderboard

Commercial Coding Models

Open-Source Coding Models

Reasoning Models 🧠

AIME 2025 Leaderboard

Reasoning Model Details

Use Cases

Multimodal Models 🎨

Leading Multimodal Models

Vision Capabilities

Audio & Video

Image Generation

Hardware Requirements 🖥️

Quick Reference by Model Size

By Hardware Tier

Quantization Explained

Comprehensive Benchmark Reference 📈

Full Benchmark Table

FrontierMath Scores

Notes on Limited Availability Models

Benchmark Glossary

Development Tools 🛠️

IDEs 💻

Agentic IDEs

Native AI Editors

VS Code Forks

Cursor Alternatives & Better Value 🎯

Pricing Reality Check (July 2026)

Value-For-Money Ranking

Official Source Basis

Why Specific Alternatives Beat Cursor

Web-Based IDEs

CLI Tools 🖥️

Autonomous Coding Agents

Assisted CLI Tools

CLI Tools by Programming Language

CLI for Programming Languages & Multiple Use

Terminal Enhancers

IDE Add-ons 🧩

Universal (Cross-Platform)

VS Code Specific

JetBrains Specific

API Providers 🔌

Model Labs (Direct)

Unified APIs & Aggregators

Model Router Services 🔀

Budget Model Aggregators

GPU Clouds

Inference Clouds

Hosted Open Source Models 🆓

Free Tier APIs

Automation 🤖

Browser Automation 🌐

Standalone AI Browsers