Clocktower Bench deception
Games · 2026-11-05
A model's Clocktower Bench rating from its games as the Demon's team, where it has to lie convincingly and coordinate in secret. It comes from the same Bradley-Terry fit and the same games as the combined rating.
Benchmark source: Clocktowerbench
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| GPT-6 Astra | 1806 | Clocktower Bench leaderboard: gpt-6-astra · Dwarkesh Patel |
| GPT-6.1 Sol | 1688 | Clocktower Bench leaderboard: gpt-6.1-sol · Dwarkesh Patel |
| Claude Fable 5 | 1643 | Clocktower Bench leaderboard: claude-fable-5 · Dwarkesh Patel |
| Claude Fable 5.1 | 1588 | Clocktower Bench leaderboard: claude-fable-5-1 · Dwarkesh Patel |
| Claude Opus 5.5 | 1585 | Clocktower Bench leaderboard: claude-opus-5-5 · Dwarkesh Patel |
| GPT-5.5 | 1560 | Clocktower Bench leaderboard: gpt-5.5 · Dwarkesh Patel |
| GPT-5.6 Sol | 1519 | Clocktower Bench leaderboard: gpt-5.6-sol · Dwarkesh Patel |
| Opus 4.6 | 1338 | Clocktower Bench leaderboard: claude-opus-4-6 · Dwarkesh Patel |
| Opus 4.8 | 1320 | Clocktower Bench leaderboard: claude-opus-4-8 · Dwarkesh Patel |
| GPT-5.4 | 1228 | Clocktower Bench leaderboard: gpt-5.4 · Dwarkesh Patel |
| Opus 4.7 | 1224 | Clocktower Bench leaderboard: claude-opus-4-7 · Dwarkesh Patel |