Best AI for fixing code and GitHub issues (Sep 2026)
As of 1 Sep 2026, the highest reported score on SWE-bench Verified is Claude Opus 5 (Anthropic) at 97.0%, per Vals AI. The source no longer updates this leaderboard. As of 1 Sep 2026, the highest reported score on LiveCodeBench is Claude Fable 5.1 (Anthropic) at 90.52%, per Vals AI. The source no longer updates this leaderboard. The leaders differ across these benchmarks, so check the one closest to your work.
Results from published AI leaderboards. Compare models within each benchmark; scores from different tests aren’t directly comparable. How we collect these results
SWE-bench Verified
Resolving real-world GitHub issues with a code patch, checked by unit tests. Run by Vals AI (benchmark by Princeton/OpenAI). Leaderboard dated 1 Sep 2026 (no longer updated by the source). Last checked 21 Oct 2026.
- Anthropic
- OpenAI
- xAI
- Meta
- DeepSeek
- Alibaba
- Z.ai
- Moonshot
- Other
Full SWE-bench Verified leaderboard: every model, all metrics and the table →
LiveCodeBench
Programming problems from LiveCodeBench, run by Vals AI. Run by Vals AI (benchmark by LiveCodeBench team). Leaderboard dated 1 Sep 2026 (no longer updated by the source). Last checked 21 Oct 2026.
- Anthropic
- OpenAI
- xAI
- DeepSeek
- Alibaba
- Moonshot
- Other
Full LiveCodeBench leaderboard: every model, all metrics and the table →
What these benchmarks tell you
- SWE-bench Verified: the share of real GitHub issues in open-source Python projects where the model’s patch makes the project’s tests pass, plus cost per issue.
- LiveCodeBench: the share of recent competitive-programming problems solved, collected after training cut-offs to limit memorisation.
What they don’t tell you
- Vals AI no longer updates either leaderboard, so models released since may be missing. For a current agentic coding signal, see Terminal-Bench 4.0 on the AI agents page.
- Python open-source fixes and contest problems are not your codebase, stack or review standards.
- The prompts below help you brief a coding agent well; they are not part of either benchmark.
Prompts for fixing code and GitHub issues
Run the same prompt on your own material with two models and compare the results.
Questions people ask
Which AI has the highest reported score for fixing code and GitHub issues?
As of 1 Sep 2026, the highest reported score on SWE-bench Verified is Claude Opus 5 (Anthropic) at 97.0%, per Vals AI. The source no longer updates this leaderboard. As of 1 Sep 2026, the highest reported score on LiveCodeBench is Claude Fable 5.1 (Anthropic) at 90.52%, per Vals AI. The source no longer updates this leaderboard. The leaders differ across these benchmarks, so check the one closest to your work.
How recent are these results?
SWE-bench Verified: leaderboard dated 1 Sep 2026, no longer updated by the source; we last checked it on 21 Oct 2026. LiveCodeBench: leaderboard dated 1 Sep 2026, no longer updated by the source; we last checked it on 21 Oct 2026.
What does the top-scoring model cost to run on these tests?
On SWE-bench Verified, Claude Opus 5 costs $1.29 (cost per test), as published by Vals AI. That is API cost on the benchmark’s own tasks, not a subscription price.
Who runs these benchmarks?
SWE-bench Verified is run by Vals AI (benchmark by Princeton/OpenAI). LiveCodeBench is run by Vals AI (benchmark by LiveCodeBench team). ChatJudge does not run the tests; we copy the published scores with a link to the source.
What don't these benchmarks tell you about fixing code and GitHub issues?
Vals AI no longer updates either leaderboard, so models released since may be missing. For a current agentic coding signal, see Terminal-Bench 4.0 on the AI agents page. Python open-source fixes and contest problems are not your codebase, stack or review standards. The prompts below help you brief a coding agent well; they are not part of either benchmark.