Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresMean human rubric score on two average-difficulty research tasks (Chess variants, Hidden messages) in Epoch's November 17, 2026 report: design and run experiments on a fixed budget and write up the findings, graded 1-4 on four equally weighted criteria.
No signal or difficulty estimate. Too few models have been measured on this benchmark to estimate its difficulty.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | thinking · Epoch Automation Reports: Claude Code; highest available reasoning setting (native value unreported); memory and skill writing disabled | 71.9 | |
| 2 | Claude Opus 5.5 | thinking · Epoch Automation Reports: Claude Code; highest available reasoning setting (native value unreported); memory and skill writing disabled | 68.8 | |
| 3 | GPT-6 Astra | thinking · Epoch Automation Reports: Codex; highest available reasoning setting (native value unreported); memory and skill writing disabled | 65.6 | |
| 4 | Claude Fable 5.1 | thinking · Epoch Automation Reports: Claude Code; highest available reasoning setting (native value unreported); memory and skill writing disabled | 62.5 | |
| 5 | GPT-6.2 Astra | thinking · Epoch Automation Reports: Codex; highest available reasoning setting (native value unreported); memory and skill writing disabled | 62.5 | |
| 6 | Gemini 4 Argon | thinking · Epoch Automation Reports: Antigravity; highest available reasoning setting (native value unreported); memory and skill writing disabled | 53.1 |