Generative AI Model Ranking Matrix

GPT-6 Astra OpenAI · 2026-09-03 · closed

OpenAI flagship. Terminal-Bench 4.0 57.7%, DeepSWE 74.1%, GPQA Diamond 96%. Text, image, and file input with tool calling, extended reasoning, and computer use. OpenAI reports OSWorld 2.0 at 72.6% and GPQA Diamond at 96.0%.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

$10 / $50 per M

Gemini 3.8 Flash Google · 2026-09-02 · closed

Cheap agent workhorse. HLE-Verified 54.9%. Price doubles on 1 Jan 2027. Text, image, audio, video, and PDF input with tool calling and extended reasoning. Google reports HLE-Verified at 54.9%.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

$0.75 / $3.75 per M

Claude Fable 5.1 Anthropic · 2026-09-01 · closed

Top of the Claude 5 family. Terminal-Bench 4.0 55.8%, DeepSWE 67.4%, HLE with tools 65%. Text, image, and file input with tool calling, extended reasoning, and computer use. Built for multi-hour agentic sessions. Cache reads cut to $0.25 per million.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

$10 / $50 per M

Claude Opus 5 Anthropic · 2026-07-24 · closed

SWE-Bench Verified 96.0 and Pro 79.2 at unchanged Opus pricing. Text and image input with tool calling, extended thinking on by default, and computer use. Anthropic reports OSWorld 2.0 at 70.6%.

SWE-Bench Verified

96%

SWE-Bench Pro

79.2%

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

$5 / $25 per M

Mistral Medium 3.5 Mistral · 2026-04-30 · open weights

128B · SWE-V 77.6 beats Devstral 2 + Qwen3.5; τ³-Telecom 91.4. Mistral Medium 3.5 supports tool calling and extended reasoning, with a 128k-token context window.

SWE-Bench Verified

77.6%

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

128k

Price in / out

$0.6 / N/A per M

Kimi K2.6 Moonshot · 2026-04-29 · open

1.1T MoE · ties GPT-5.5 on SWE-Pro at $0.6/M; HLE 54 leads all. Kimi K2.6 supports vision input, tool calling and extended reasoning, with a 256k-token context window.

SWE-Bench Verified

80.2%

SWE-Bench Pro

58.6%

Terminal-Bench

66.7%

AA Index

N/A

Context

256k

Price in / out

$0.6 / N/A per M

MiMo-V2.5-Pro Xiaomi · 2026-04-29 · open

1.02T MoE / 42B active · matches Opus 4.6 on SWE-Pro w/ 40% fewer tokens. MiMo-V2.5-Pro supports tool calling and extended reasoning, with a 1M-token context window.

SWE-Bench Verified

N/A

SWE-Bench Pro

57.2%

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

N/A / N/A per M

Gemma 4 31B-it Google · 2026-04-29 · open

31B dense · AIME 89.2, GPQA 84.3, τ²-Retail 86.4 (12× jump from G3). Gemma 4 31B-it supports vision input and tool calling, with a 256k-token context window.

SWE-Bench Verified

52%

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

256k

Price in / out

N/A / N/A per M

Nemotron-3-Nano-Omni NVIDIA · 2026-04-29 · open

Fresh (~17h) · 30B-A3B any-to-any reasoning. Nemotron-3-Nano-Omni supports vision input, audio, tool calling and extended reasoning, with a 128k-token context window.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

128k

Price in / out

N/A / N/A per M

DeepSeek V4-Pro DeepSeek · 2026-04-27 · open

1.6T MoE / 49B active · matches Opus 4.6 at fraction of cost. DeepSeek V4-Pro supports tool calling and extended reasoning, with a 128k-token context window.

SWE-Bench Verified

80.6%

SWE-Bench Pro

55.4%

Terminal-Bench

67.9%

AA Index

N/A

Context

128k

Price in / out

$0.9 / N/A per M

DeepSeek V4-Flash DeepSeek · 2026-04-27 · open

Fresh · 158B fast variant of V4. DeepSeek V4-Flash supports tool calling and extended reasoning, with a 128k-token context window.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

128k

Price in / out

$0.3 / N/A per M

GPT-5.5 OpenAI · 2026-04-24 · closed

SOTA Terminal-Bench 82.7; GDPval 84.9; OSWorld 78.7; ties Opus 4.7 on SWE-V. GPT-5.5 supports vision input, audio, tool calling, extended reasoning and computer use, with a 1M-token context window.

SWE-Bench Verified

87.6%

SWE-Bench Pro

58.6%

Terminal-Bench

82.7%

AA Index

N/A

Context

1M

Price in / out

$5 / $30 per M

Qwen3.5-397B-A17B Alibaba · 2026-04-24 · open

403B MoE / 17B active · GPQA 88.4, MMLU-Pro 87.8, SWE-V 80.0. Qwen3.5-397B-A17B supports vision input, tool calling and extended reasoning, with a 256k-token context window.

SWE-Bench Verified

80%

SWE-Bench Pro

N/A

Terminal-Bench

54%

AA Index

N/A

Context

256k

Price in / out

N/A / N/A per M

Qwen3.6-35B-A3B Alibaba · 2026-04-24 · open

36B / 3B active · 73.4 SWE-V on tiny active params. Qwen3.6-35B-A3B supports vision input, tool calling and extended reasoning, with a 256k-token context window.

SWE-Bench Verified

73.4%

SWE-Bench Pro

49.5%

Terminal-Bench

N/A

AA Index

N/A

Context

256k

Price in / out

N/A / N/A per M

Tencent Hy3-preview Tencent · 2026-04-24 · open (preview)

295B / 21B active · +40pts SWE-V vs Hy2; topped Tsinghua math PhD exam. Tencent Hy3-preview supports vision input, tool calling and extended reasoning, with a 128k-token context window.

SWE-Bench Verified

74.4%

SWE-Bench Pro

N/A

Terminal-Bench

54.4%

AA Index

N/A

Context

128k

Price in / out

N/A / N/A per M

Qwen3.6-Max-Preview Alibaba · 2026-04-20 · closed

Agent loops: preserve_thinking across tool calls. Qwen3.6-Max-Preview supports vision input, tool calling and extended reasoning, with a 1M-token context window.

SWE-Bench Verified

79.5%

SWE-Bench Pro

N/A

Terminal-Bench

77.1%

AA Index

N/A

Context

1M

Price in / out

$6 / N/A per M

MiniMax M2.7 MiniMax · 2026-04-20 · open

229B · self-evolving; matches Codex on SWE-Pro. MiniMax M2.7 supports tool calling and extended reasoning, with a 205k-token context window.

SWE-Bench Verified

78%

SWE-Bench Pro

56.22%

Terminal-Bench

57%

AA Index

N/A

Context

205k

Price in / out

$0.3 / N/A per M

Claude Opus 4.7 Anthropic · 2026-04-16 · closed

Leads SWE-Bench Verified (87.6) and Pro (64.3); new tokenizer (~35% more tokens). Claude Opus 4.7 supports vision input, tool calling, extended reasoning and computer use, with a 1M-token context window.

SWE-Bench Verified

87.6%

SWE-Bench Pro

64.3%

Terminal-Bench

69.4%

AA Index

N/A

Context

1M

Price in / out

$5 / $25 per M

GLM-5.1 Zhipu / Z.ai · 2026-04-07 · open (MIT)

754B · #1 SWE-Bench Pro among open weights. GLM-5.1 supports vision input, tool calling and extended reasoning, with a 128k-token context window.

SWE-Bench Verified

79%

SWE-Bench Pro

58.4%

Terminal-Bench

N/A

AA Index

N/A

Context

128k

Price in / out

$0.11 / N/A per M

Grok 4.20 Beta 2 xAI · 2026-03-03 · closed

4-agent backbone · IFBench #1 (83); 8% of Opus cost. Grok 4.20 Beta 2 supports vision input and tool calling, with a 256k-token context window.

SWE-Bench Verified

75%

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

256k

Price in / out

$2 / N/A per M

GPT-5.3 Codex OpenAI · 2026-03 · closed

Full SDLC agent: debug, terminal, PRDs, tests. GPT-5.3 Codex supports vision input, tool calling, extended reasoning and computer use, with a 400k-token context window.

SWE-Bench Verified

80%

SWE-Bench Pro

56.8%

Terminal-Bench

77.3%

AA Index

N/A

Context

400k

Price in / out

$10 / N/A per M

Gemini 3.1 Pro Google · 2026-03 · closed

Adjustable Deep Think; agentic browsing (BrowseComp 85.9). Gemini 3.1 Pro supports vision input, audio, tool calling, extended reasoning and computer use, with a 2M-token context window.

SWE-Bench Verified

78%

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

2M

Price in / out

$7 / N/A per M

Devstral-2-123B Mistral · 2026-02-25 · open weights

125B · code-specialized variant. Devstral-2-123B supports tool calling and computer use, with a 128k-token context window.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

128k

Price in / out

$0.6 / N/A per M

Claude Opus 4.6 Anthropic · 2026-02 · closed

Large-codebase reasoning, multi-file refactors. Claude Opus 4.6 supports vision input, tool calling, extended reasoning and computer use, with a 1M-token context window.

SWE-Bench Verified

80.8%

SWE-Bench Pro

57.3%

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

$15 / N/A per M

Mistral Large 3 Mistral · 2025-12-02 · open (Apache 2.0)

675B MoE / 41B active · non-reasoning (AIME ~40, GPQA ~44) · Apache 2.0. Mistral Large 3 supports vision input and tool calling, with a 256k-token context window.

SWE-Bench Verified

N/A

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

256k

Price in / out

$2 / N/A per M

Claude Sonnet 4.6 Anthropic · 2025-12 · closed

Default coding driver. 98% of Opus at ⅕ cost. Claude Sonnet 4.6 supports vision input, tool calling, extended reasoning and computer use, with a 1M-token context window.

SWE-Bench Verified

79.6%

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

1M

Price in / out

$3 / $15 per M

Claude Haiku 4.5 Anthropic · 2025-10-15 · closed

4 to 5 times faster than Sonnet 4.5; cheap multi-agent driver. Claude Haiku 4.5 supports vision input, tool calling, extended reasoning and computer use, with a 200k-token context window.

SWE-Bench Verified

73.3%

SWE-Bench Pro

N/A

Terminal-Bench

N/A

AA Index

N/A

Context

200k

Price in / out

$1 / $5 per M

gpt-astra-latest OpenAI · 2026-09-11 · closed · live

Insufficient information to confirm frontier-class capabilities; model card provides no substantive details beyond routing to a model family, making full evaluation impossible.

AA Index

Awaiting benchmarks

Context

1.1M

Price in / out

$10 / $50 per M

Found via

OpenRouter

gpt-sol-latest OpenAI · 2026-09-11 · closed · live

Insufficient information to assess the model's true capabilities, architecture, or scale; the model card only states it redirects to the latest GPT Sol family model with no further details.

AA Index

Awaiting benchmarks

Context

1.1M

Price in / out

$2 / $10 per M

Found via

OpenRouter

gpt-terra-latest OpenAI · 2026-09-11 · closed · live

Insufficient information to assess the model's true capabilities, architecture, or scale; the model card provides only a routing description with no technical details.

AA Index

Awaiting benchmarks

Context

1.1M

Price in / out

$2 / $12 per M

Found via

OpenRouter

gpt-luna-latest OpenAI · 2026-09-11 · closed · live

Insufficient information to determine if this is a frontier-class model; the model card provides almost no technical details, params are unknown, and 'GPT Luna' is not a recognized public OpenAI release as of the knowledge cutoff.

AA Index

Awaiting benchmarks

Context

1.1M

Price in / out

$0.19999999999999998 / $1.2 per M

Found via

OpenRouter

Nex-N2.5-Pro nex-agi · 2026-09-11 · open · live

Nex-N2.5-Pro is a ~397B MoE agentic model with multimodal (image+text) capabilities, computer use, and web browsing, representing a meaningful frontier-class post-training effort on a large-scale foundation.

AA Index

Awaiting benchmarks

Context

N/A

Price in / out

N/A / N/A per M

Found via

Hugging Face

V4.1-Flash DeepSeek · 2026-09-10 · open · live

Novel Causal Encoder-Decoder (CED) architecture from a major lab is significant, but small active parameter count (8B) limits scale score despite the architectural innovation.

AA Index

Awaiting benchmarks

Context

1M

Price in / out

$0.15 / $0.6 per M

Found via

OpenRouter

gpt-6-astra-pro OpenAI · 2026-09-04 · closed · live

GPT-6 Astra Pro appears to be a frontier-class model from OpenAI with a very large context window and high pricing suggesting significant scale, but it is merely a reasoning-mode variant of GPT-6 Astra rather than a distinct model, and no verifiable metadata (params, architecture, capabilities) is available to confirm claims.

AA Index

Awaiting benchmarks

Context

1.1M

Price in / out

$10 / $50 per M

Found via

OpenRouter

GLM-5.3 Zhipu / Z.ai · 2026-09-04 · open · live

GLM-5.3 is a 753B-parameter open-weights MoE model achieving frontier-class coding and cyber capability benchmarks, representing a significant post-training advancement over GLM-5.2 with SOTA results on multiple public benchmarks.

AA Index

Awaiting benchmarks

Context

N/A

Price in / out

N/A / N/A per M

Found via

Hugging Face

qwen3.8-max-0902 Alibaba · 2026-09-03 · closed · live

2.4T parameter MoE model with multimodal capabilities (text/image/video) and 1M context window represents frontier-class scale and capability.

AA Index

Awaiting benchmarks

Context

1M

Price in / out

$2 / $6 per M

Found via

OpenRouter

Hy4-preview Tencent · 2026-08-28 · open · live

Tencent's 780B MoE flagship model represents frontier-class scale and capability with a new-generation architecture from a major lab.

AA Index

Awaiting benchmarks

Context

N/A

Price in / out

N/A / N/A per M

Found via

Hugging Face

Qwen3.8-Flash-Next Alibaba · 2026-08-26 · open · live

Qwen3.8 Flash appears to be a capable multimodal reasoning model with a 1M context window and broad capabilities, but parameter count is unknown and it may be a smaller/flash variant, keeping composite below include threshold.

AA Index

Awaiting benchmarks

Context

1M

Price in / out

$0.15 / $0.47 per M

Found via

OpenRouter

GLM-5.3-Flash Zhipu / Z.ai · 2026-08-26 · open · live

GLM-5.3-Flash features a novel hybrid sparse/linear attention architecture with a very large 1.3M token context window suited for agent tasks, but parameter count is unknown and it's a flash/efficient variant rather than a frontier-scale model.

AA Index

Awaiting benchmarks

Context

1.3M

Price in / out

$0.15 / $0.5 per M

Found via

OpenRouter

V4-Flash-0731 DeepSeek · 2026-08-01 · open · live

DeepSeek-V4-Flash-0731 is a frontier-class 304B-parameter MoE model from a major lab with a dedicated technical report, high download counts, and represents a new official release superseding a preview version.

AA Index

Awaiting benchmarks

Context

N/A

Price in / out

N/A / N/A per M

Found via

Hugging Face