Epoch Automation Reports
Agents · 2026-10-08
Human rubric evaluation of six native agent systems on 11 tasks in five categories, published October 8, 2026. Overall is the unweighted mean of category scores, not of individual tasks. Task scores use weighted rubric sections where provided and deduct human-intervention points. Highest available reasoning was selected, but exact native effort values were not disclosed; memory and skill writing were disabled.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 65.3 |
| GPT-6 Astra | 65.1 |
| Grok 4.6 | 59.1 |
| Kimi K3 | 52.4 |
| Gemini 3.8 Flash | 42.3 |