// BENCHMARK

SWE-bench Verified benchmark

AI model leaderboard for the SWE-bench Verified benchmark. Compare how large language models score on SWE-bench Verified, see the full ranking, and understand what this AI benchmark measures. Claude Fable 5 currently leads with 95%. Human-verified subset of SWE-bench filtered for unambiguous task descriptions and reliable test suites.

Leaderboard

# Model Organization Score Variant Source
#1 Claude Fable 5 Anthropic 95% official ↗
#2 Claude Mythos 5 Anthropic 95% shared with Claude Fable 5 (same underlying model) official ↗
#3 Claude Opus 4.7 Anthropic 82.4% agentic official ↗
#4 Claude Opus 4.5 Anthropic 80.9% official ↗
#5 Claude Opus 4.6 Anthropic 80.8% agentic official ↗
#6 MiniMax M3 MiniMax 80.5%
#7 Qwen3.6-Max-Preview Qwen 80.5% official ↗
#8 GPT-5.5 Pro OpenAI 80.5% parallel-compute official ↗
#9 Kimi K2.6 Moonshot 80.2% agentic official ↗
#10 DeepSeek V4-Pro DeepSeek 79.4% thinking official ↗
#11 Claude Sonnet 4.6 Anthropic 78.2% agentic official ↗
#12 Mistral Medium 3.5 Mistral AI 77.6% official ↗
#13 Claude Sonnet 4.5 Anthropic 77.2% cited-devstral-readme official ↗
#14 Claude Sonnet 4.5 Anthropic 77.2% cited-deepseek-v3-2 official ↗
#15 Qwen3.6-Plus Qwen 76.9% agentic official ↗
#16 Trinity-Large-Thinking Arcee AI 76.5% official ↗
#17 Gemini 3 Pro Google DeepMind 76.2% cited-deepseek-v3-2 official ↗
#18 Gemini 3.1 Pro Google DeepMind 76.2% official ↗
#19 Gemini 3 Pro Google DeepMind 76.2% cited-devstral-readme official ↗
#20 GPT-5.5 OpenAI 76% agentic official ↗
#21 GLM-5.1 Zhipu AI 75.6% official ↗
#22 Kimi K2.5 Moonshot 75.6% official ↗
#23 GPT-5.4 Thinking OpenAI 74.9% thinking official ↗
#24 GPT-5 OpenAI 74.9% cited-deepseek-v3-2 official ↗
#25 MAI-Thinking-1 Microsoft 73.5% official ↗
#26 MiniMax M2.7 MiniMax 73.4% official ↗
#27 DeepSeek-V3.2 DeepSeek 73.1% thinking official ↗
#28 DeepSeek-V3.2 DeepSeek 73.1% cited-devstral-readme official ↗
#29 GLM-5 Zhipu AI 72.8% official ↗
#30 Devstral 2 Mistral AI 72.2% from-hf-readme official ↗
#31 DeepSeek V4-Flash DeepSeek 72.1% thinking official ↗
#32 GPT-5.4 OpenAI 72% official ↗
#33 Kimi K2 Thinking Moonshot 71.3% cited-devstral-readme official ↗
#34 Kimi K2 Thinking Moonshot 71.3% thinking-w-tools official ↗
#35 Kimi K2 Thinking Moonshot 71.3% thinking-cited-deepseek-v3-2 official ↗
#36 Grok Build 0.1 xAI 70.8% official ↗
#37 NVIDIA Nemotron 3 Ultra NVIDIA 70.7% official ↗
#38 Claude Sonnet 4 Anthropic 70.4% official ↗
#39 MiniMax-M2 MiniMax 69.4% cited-deepseek-v3-2 official ↗
#40 MiniMax-M2 MiniMax 69.4% cited-devstral-readme official ↗
#41 Kimi K2 Moonshot 69.2% k2-0905-w-tools-cited-k2-thinking official ↗
#42 o3 OpenAI 69.1% official ↗
#43 Devstral Small 2 Mistral AI 68% from-hf-readme official ↗
#44 GLM 4.6 Zhipu AI 68% cited-devstral-readme official ↗
#45 DeepSeek-V3.2-Exp DeepSeek 67.8% from-hf-readme official ↗
#46 Claude Opus 4 Anthropic 67.8% official ↗
#47 North Mini Code 1.0 Cohere 67.6%
#48 Kimi K2 Moonshot 65.8% official ↗
#49 Kimi K2 Moonshot 65.4% official ↗
#50 GLM 4.5 Zhipu AI 64.2% official ↗
#51 Intern-S2-Preview Shanghai AI Laboratory 64% official ↗
#52 gpt-oss-120b OpenAI 62.4% high reasoning official ↗
#53 gpt-oss-120b OpenAI 62.4% high official ↗
#54 gpt-oss-20b OpenAI 60.7% high official ↗
#55 gpt-oss-20b OpenAI 60.7% high reasoning official ↗
#56 LongCat-Flash Meituan Inc 60.4% official ↗
#57 GLM-4.5-Air Zhipu AI 57.6% official ↗
#58 MiniMax-M1-80k MiniMax 56% Agentless official ↗
#59 MiniMax-M1-40k MiniMax 55.6% Agentless official ↗
#60 gemini-2.5-pro Google DeepMind 49% official ↗
#61 GPT-4.1 OpenAI 48.6% official ↗
#62 DeepSeek-R1-0528 DeepSeek 41.4% official ↗
#63 DeepSeek-V2.5 DeepSeek 16.8% from-hf-readme official ↗

Frequently asked questions about SWE-bench Verified

What is the SWE-bench Verified benchmark?

Human-verified subset of SWE-bench filtered for unambiguous task descriptions and reliable test suites.

How is the SWE-bench Verified benchmark scored?

SWE-bench Verified is scored using the resolved (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on SWE-bench Verified?

As of the latest reported scores on GenAIList, Claude Fable 5 achieves the highest result on SWE-bench Verified with a score of 95%.

Is a higher SWE-bench Verified score better?

Yes. On SWE-bench Verified a higher score indicates better performance, so models near the top of the leaderboard are the strongest.