EBR-bench v2
Games · 2026-07-27
EBR-bench v2 topline averages each independent sample's best objective-completion fraction from its final two playthroughs, using at least ten samples per starting state. It is an averaged run-level score over 21 possible objectives, not one binary trial per objective, and is kept separate from v1's final-20%-mean protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 45.7 |
| Claude Fable 5 | 39.5 |
| GPT-5.6 Sol | 38.6 |
| Opus 4.8 | 25.2 |
| GPT-5.5 | 21.0 |