Evo-Bench - Office: leaderboard
Evo-Bench (Huang et al., 2026) measures harness evolution: within 20 iterations the evaluated model rewrites a minimal CodeAct agent harness from validation feedback, and a fixed DeepSeek-V4-Flash policy then runs the frozen harness on held-out tasks. This board averages the held-out GDPval and APEX-Agents office tasks, in percent; higher is better.
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 64 GDPval and 64 APEX-Agents held-out tasks; one run). Source: arxiv.org. Saturation forecast: Around 2030. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 41.6 | #29 (GPT-5.6 Sol) |
| 2 | Gemma 4 31B (IT) (Thinking) | 40.4 | #343 (Gemma 4 31B (IT)) |
| 3 | Claude Opus 4.8 (Max) | 39.7 | #56 (Claude Opus 4.8) |
| 4 | GLM-5.2 (Max) | 39.2 | #137 (GLM-5.2) |
| 5 | DeepSeek V4 Pro (Max) | 39.1 | #181 (DeepSeek V4 Pro) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evo-bench-office · How It Works · Data refreshed daily, snapshot 2026-11-12.