SWE-bench Leaderboard

Every major AI model ranked by SWE-bench — the benchmark that measures whether a model can resolve real GitHub issues, not toy problems. Pricing and context columns are included because a leaderboard you can't act on is just trivia.

Scores from provider publications and public leaderboards · Pricing verified daily

Current leader

GPT-5.6 Sol96.2%OpenAI · $2/1M input · 1.1M tokens context

Best value in the top 10: Gemini 3.7 Flash — 80.8% at $0.75/1M input tokens.

GPT-5.6 Sol vs Claude Opus 5: full comparison →Cheapest above 70%: Muse Glimmer 30B ($0.35/1M) →What's a good score? →

#	Model	Provider	SWE-bench	Input $/1M	Context
1	GPT-5.6 Sol	OpenAI	96.2%	$2	1.1M tokens
2	Claude Opus 5	Anthropic	96%	$5	1M tokens
3	Claude Mythos 5	Anthropic	95.5%	$10	1M tokens
4	Claude Fable 5	Anthropic	95%	$10	1M tokens
5	Claude Opus 4.8	Anthropic	88.6%	$5	1M tokens
6	Claude Opus 4.7	Anthropic	87.6%	$5	1M tokens
7	Grok 4.5	xAI	86.6%	$2	500k tokens
8	Claude Sonnet 5	Anthropic	85.2%	$2	1M tokens
9	Claude Opus 4.6	Anthropic	80.8%	$5	1M tokens
10	Gemini 3.7 Flash	Google	80.8%	$0.75	1.0M tokens
11	Gemini 3.1 Pro	Google	80.6%	$2	2M tokens
12	Qwen 3.7 Max	Alibaba	80.4%	$2.5	1M tokens
13	GPT-5.2	OpenAI	80%	$1.75	200k tokens
14	Claude Sonnet 4.6	Anthropic	79.6%	$3	1M tokens
15	Mistral Medium 3.5	Mistral	77.6%	$1.5	256k tokens
16	Muse Spark	Meta	77.4%	$1.25	1.0M tokens
17	Muse Glimmer 30B	Meta	76%	$0.35	131k tokens
18	GPT-5.4	OpenAI	74.9%	$2.5	272k tokens
19	Grok 4	xAI	54%	$2	2M tokens
20	DeepSeek R1	DeepSeek	49.2%	$0.55	128k tokens
21	GPT-4o	OpenAI	46%	$2.5	128k tokens
22	Claude Haiku 4	Anthropic	43%	$0.8	200k tokens
23	DeepSeek V3	DeepSeek	42%	$0.27	128k tokens
24	Gemini 3.1 Flash	Google	35%	$0.5	1M tokens
25	Llama 4 Maverick	Meta	32%	$0.6	256k tokens
26	Mistral Large 2	Mistral	28%	$3	128k tokens
27	GPT-4o Mini	OpenAI	23.6%	$0.15	128k tokens

All benchmark scores →Best AI for coding →GPT-5.6 Sol alternatives →API cost calculator →What is a good SWE-bench score? →Verified vs full SWE-bench →

FAQ

What is SWE-bench?

SWE-bench is a benchmark that tests whether an AI model can resolve real GitHub issues from real open-source repositories — write a patch, pass the tests. Unlike puzzle-style benchmarks, it measures the messy, multi-file work software engineers actually do, which is why it has become the standard for comparing coding models.

Which AI model has the highest SWE-bench score?

GPT-5.6 Sol (OpenAI) currently leads at 96.2%, ahead of Claude Opus 5 at 96%.

What is a good SWE-bench score?

Anything above 70% is frontier-class in 2026 — the model can resolve most real GitHub issues autonomously. The current leader, GPT-5.6 Sol, is at 96.2%. Two years ago the best models scored under 20%, which is how fast this benchmark moves.

Does the highest SWE-bench score mean the best coding model for me?

Not always. Score-per-dollar matters for daily work: Gemini 3.7 Flash delivers 80.8% at $0.75/1M input tokens, which is the best value in the top 10. Reserve the outright leader for the hardest tasks and route volume work to the value pick.