Evo-Bench - General: leaderboard

Metric: Held-out score (%; the fixed DeepSeek-V4-Flash policy runs the harness the model evolved from a CodeAct seed within 20 iterations, 1,000 steps and 48 hours; 64 held-out Claw-Eval tasks scored by Pass^3; one run). Source: arxiv.org. Saturation forecast: Around June 2028. 9 models tracked.

Top models

#ModelScore
1GPT-5.6 Sol (Max)59.4
2Claude Opus 4.8 (Max)56.3
3Gemma 4 31B (IT) (Thinking)50
4GLM-5.2 (Max)48.4
5DeepSeek V4 Pro (Max)48.4

Interactive version: theaggregate.ai/benchmark?slug=evo-bench-general · How It Works · Data refreshed daily, snapshot 2026-09-26.