Merge-Bench (All Languages) - Normalized Equivalence: leaderboard

Merge-Bench (2026) gives models real merge-conflict hunks from open-source repositories in 11 programming languages. This board counts resolutions that match the developer resolution once formatting and comments are normalized away, in percent; higher is better.

Metric: Code-normalized equivalent (%): share of conflicts whose output matches the developer resolution after removing formatting and comments, all 7,938 merge-conflict hunks from 1,439 repositories in 11 languages, the model returns the resolved snippet in Markdown, compared with the resolution the developers committed; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro52.6#276
2O3 Pro45.1#209
3Claude Opus 444.8#326
4DeepSeek R1 052836.5#428
5Grok 431.7#319
6Qwen 3 235B A22B30.6#628

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=merge-bench-all-languages-normalized-equivalence · How It Works · Data refreshed daily, snapshot 2026-11-21.