Evo-Bench - Search: 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 128 BrowseComp and 128 HLE held-out tasks; one run). Source: arxiv.org. Saturation forecast: Around May 2028. 9 models tracked.

Top models

#ModelScore
1Claude Opus 4.8 (Max)46.5
2GLM-5.2 (Max)45.4
3GPT-5.6 Sol (Max)44.5
4DeepSeek V4 Pro (Max)34.4
5Gemma 4 31B (IT) (Thinking)24.2

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