Epoch AI - Dtbench: leaderboard
Newcomb-like problems, where what an agent decides is evidence about what copies of it will decide. 537 handcrafted multiple-choice questions from Redwood Research; Epoch scores only the 407 with objectively correct answers and drops the 130 that probe attitudes instead. Random guessing gets 40 and the authors put the ceiling at or near 100.
Metric: Score. Source: epoch.ai. Saturation forecast: Around June 2027. 235 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 98.93 | #3 (Claude Opus 5.5) |
| 2 | Claude Fable 5 (Max) | 98.4 | #16 (Claude Fable 5) |
| 3 | Claude Opus 5 (High) | 97.87 | #12 (Claude Opus 5) |
| 4 | Claude Opus 5 (xHigh) | 97.87 | #12 (Claude Opus 5) |
| 5 | Claude Opus 5 (Max) | 97.6 | #12 (Claude Opus 5) |
| 6 | Claude Fable 5.1 (xHigh) | 97.6 | #9 (Claude Fable 5.1) |
| 7 | GPT-6 Sol (Max) | 97.33 | #14 (GPT-6 Sol) |
| 8 | GPT-6 Astra (xHigh) | 97.33 | #6 (GPT-6 Astra) |
| 9 | Gemini 4 Argon (High) | 97.33 | #8 (Gemini 4 Argon) |
| 10 | Grok 4.6 (xHigh) | 97.33 | #26 (Grok 4.6) |
| 11 | Claude Sonnet 5.5 (Max) | 97.33 | #11 (Claude Sonnet 5.5) |
| 12 | Claude Opus 5.5 (xHigh) | 97.33 | #3 (Claude Opus 5.5) |
| 13 | Claude Fable 5 (xHigh) | 97.33 | #16 (Claude Fable 5) |
| 14 | GPT-6 Astra | 97.07 | #6 |
| 15 | GPT-6 Astra (Medium) | 97.07 | #6 (GPT-6 Astra) |
No result here: #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=epoch-ai-dtbench · How It Works · Data refreshed daily, snapshot 2026-10-20.