Clocktower Bench - Deception: leaderboard
Clocktower Bench (Dwarkesh Patel, 2026) evaluates strategic social deduction in Blood on the Clocktower: one model plays all good seats, another all evil seats, with isolated seat contexts and side-swapped deals. The published 386-game study pools the authors' v16/v17 protocols over 11 models, each at high effort and an 8,000-token per-turn cap, and reports deception Bradley–Terry Elo, centered at 1500. Detection measures good-town play; deception measures evil-team play. Ratings use the authors' setup adjustments and 90% side-swapped-pair bootstrap intervals. Retry-bug originals and mixed-team ablations are excluded from this fit. Higher is better; Elo has no theoretical bounds.
Metric: Native Bradley–Terry Elo rating for evil-team deception; 1500 centers the rated model population; higher is better; unbounded. Source: clocktowerbench.com. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1806 | |
| 2 | GPT-6.1 Sol (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1688 | |
| 3 | Claude Fable 5 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1643 | |
| 4 | Claude Fable 5.1 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1588 | |
| 5 | Claude Opus 5.5 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1585 | |
| 6 | GPT-5.5 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1560 | |
| 7 | GPT-5.6 Sol (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1519 | |
| 8 | Claude Opus 4.6 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1338 | |
| 9 | Claude Opus 4.8 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1320 | |
| 10 | GPT-5.4 (Clocktower Run: study=2026-v16-v17; effort=high; max_tokens=8000) | 1228 |
Interactive version: theaggregate.ai/benchmark?slug=clocktower-bench-deception · How It Works · Data refreshed daily, snapshot 2026-11-06.