Coding

SWE-bench Verified: 2026 AI Leaderboard

Fix real GitHub issues in 12 open-source Python repos.

What it tests

SWE-bench Verified is a 500-issue subset of SWE-bench that has been human-validated as solvable. Each task is a real Python GitHub issue; the model is given the repo, the issue, and must produce a patch that makes the project's test suite pass.

How it is scored

Percentage of issues where the generated patch passes all hidden tests. This is end-to-end agentic coding, not just code-completion. Scores above 70% are state-of-the-art; a year ago it was 30%.

Why it matters

SWE-bench Verified is the closest industry-standard benchmark to 'can this model actually do my job'. It rewards code-reading, multi-file editing, and test-driven iteration -- not just autocomplete.

Leaderboard (9 models)

Sorted by SWE-bench Verifiedscore. Tier column shows the tool's overall AIToolTier rank, which blends this benchmark with pricing, features, and real-world usability.

#ModelTierSWE-bench Verified score
1Claude (Anthropic)
Claude Opus 4.7 (4.6 baseline scores shown; 4.7 announced 13% coding lift, 3x production task completion)
A80.8%
2Gemini (Google)
Gemini 3.1 Ultra
A80.6%
3MiniMax M2 / M2.5
MiniMax M2.5 (230B/10B active MoE)
A80.2%
4Kimi K2.5 (Moonshot)
Kimi K2.5 (1T/32B active MoE)
A78.5%
5Codex (OpenAI)
GPT-5.3-Codex
A72%
6ChatGPT
GPT-5.4
A72%
7Qwen (Alibaba)
Qwen3.5-397B MoE
A69.4%
8DeepSeek
DeepSeek V3.2
A67.8%
9GLM / Z.ai (Zhipu AI)
GLM-5.1 (744B MoE / 40B active)
A64.2%

About SWE-bench Verified

Creator
Princeton & OpenAI, 2023 (Verified subset 2024)
Unit
% (max 100)

Other benchmarks