Evo-Bench: 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; mean of the five source benchmarks (BrowseComp, HLE, GDPval, APEX-Agents, Claw-Eval) over the 448-task evaluation suite; one run). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.

Top models

#ModelScore
1GPT-5.6 Sol (Max)46.3
2Claude Opus 4.8 (Max)45.8
3GLM-5.2 (Max)43.5
4DeepSeek V4 Pro (Max)39.1
5Gemma 4 31B (IT) (Thinking)35.9

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