SWE-bench Verified

Real-world code issue resolution.

Get in touch Read research

500Public tasks

93.5%Highest score

The SWE-bench Verified leaderboard

Tests whether LLMs can resolve real-world GitHub issues by editing codebases, spanning 2,294 problems from 12 popular Python repos and requiring multi-file, long-context changes.

Paper Data Code Sample task

Model

Score

Fable 5Max

93.5%

Opus 5Max

93.5%

Grok 4.5High

81.7%

Gemini 3.6 FlashHigh

81.1%

GPT-5.5xHigh

78.8%

60%

70%

80%

90%

100%

SWE-bench Verified

The SWE-bench Verified leaderboard

The latest on frontier AI performance, straight to your inbox.