FrontierCode

Hard real-world coding tasks run in agent harnesses (Cognition).

Published by

Cognitiondata via Epoch AI Benchmarking Hub

Category

Coding

Index weight

0.5

Models

29

Data as of

9 Sept 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

1. 1Claude Fable 553.5%
2. 2Claude Opus 553.4%
3. 3GPT-6 Astra53.3%
4. 4Claude Fable 5.150.9%
5. 5Grok 4.648.0%
6. 6GPT-5.6 Sol47.5%
7. 7Claude Opus 4.846.5%
8. 8Kimi K344.2%
9. 9Gemini 3.7 Flash43.6%
10. 10GPT-5.543.0%
11. 11Claude Sonnet 542.7%
12. 12Grok 4.542.4%
13. 13SWE-1.742.0%
14. 14GPT-5.6 Terra41.3%
15. 15GPT-5.6 Luna39.8%

SearchAll providersAlibabaAnthropicCognitionCursorDeepSeekGoogleMiniMaxMistral AIMoonshot AIOpenAIxAIZhipu AIOpen weights29 of 29

#	Model	Score ↓	Run ↕	Index ↕	Published ↕
1	Claude Fable 5Anthropic	53.5%	–	70.4	–
2	Claude Opus 5Anthropic	53.4%	–	66.8	–
3	GPT-6 AstraOpenAI	53.3%	–	70.6	–
4	Claude Fable 5.1Anthropic	50.9%	–	68.0	–
5	Grok 4.6xAI	48.0%	–	63.6	–
6	GPT-5.6 SolOpenAI	47.5%	–	65.3	–
7	Claude Opus 4.8Anthropic	46.5%	–	64.7	–
8	Kimi K3Moonshot AIopen	44.2%	–	66.5	–
9	Gemini 3.7 FlashGoogle	43.6%	–	61.7	–
10	GPT-5.5OpenAI	43.0%	–	66.6	–
11	Claude Sonnet 5Anthropic	42.7%	–	61.2	–
12	Grok 4.5xAI	42.4%	–	63.4	–
13	CSWE-1.7Cognition	42.0%	–	–	–
14	GPT-5.6 TerraOpenAI	41.3%	–	62.3	–
15	GPT-5.6 LunaOpenAI	39.8%	–	56.6	–
16	Claude Opus 4.7Anthropic	38.5%	–	65.5	–
17	Gemini 3.6 FlashGoogle	34.4%	–	60.7	–
18	Kimi K2.7 CodeMoonshot AIopen	30.1%	–	56.2	–
19	GPT-5.4 miniOpenAI	27.0%	–	54.8	–
20	Claude Opus 4.6Anthropic	26.6%	–	63.1	–
21	CComposer 2.5Cursor	25.6%	–	–	–
22	GLM-5.2Zhipu AIopen	24.5%	–	50.3	–
23	Claude Sonnet 4.6Anthropic	24.3%	–	60.3	–
24	DeepSeek V4 FlashDeepSeekopen	18.8%	0731, via chisel	59.1	–
25	DeepSeek V4 ProDeepSeekopen	17.6%	–	48.8	–
26	MiniMax-M3MiniMaxopen	14.7%	–	56.9	–
27	Qwen3.7 PlusAlibaba	10.2%	–	59.7	–
28	CSWE-1.6Cognition	9.4%	–	–	–
29	Mistral Medium (latest)Mistral AI	8.0%	medium, 2604, via chisel	45.8	–