# Clocktower Bench (Blood on the Clocktower) (Clocktower Bench)

> Frontier models play the social deduction game Blood on the Clocktower against each other, one model as the town and one as the Demon's team. The rating combines catching liars (detection) with lying well (deception).

Canonical page: https://benchlm.ai/benchmarks/clocktowerbench

- Category: [Reasoning](/reasoning)
- Last updated: November 6, 2026

## About Clocktower Bench

- Year: 2026
- Tasks: 386 rated head-to-head games of Blood on the Clocktower (10 to 12 players)
- Format: Relative rating from side-swapped game pairs
- Difficulty: Social deduction, deception and multi-agent reasoning
- Paper: [Clocktower Bench leaderboard](https://clocktowerbench.com/)

Each game pits one model playing every good seat against another playing every evil seat, and each deal is replayed with the sides swapped. A single Bradley-Terry fit predicts the winner from both models' ratings, a good-side advantage and the setup (which Minions are in play; 10, 11 or 12 players). Ratings are in Elo points, centred so that 1500 is the average of the 11 rated models, with 90% intervals from a bootstrap over side-swapped pairs. The source also publishes separate detection (as good) and deception (as evil) ratings. Every model plays at its lab's high reasoning effort with an 8,000-token ceiling per turn. The rated set is the 2026 study: 386 games across 11 OpenAI and Anthropic models released in 2026.

Clocktower Bench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (11 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 1859 |
| 2 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 1734 |
| 3 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 1595 |
| 4 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 1591 |
| 5 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 1549 |
| 6 | [Claude Fable 5](/models/claude-fable) | Anthropic | 1544 |
| 7 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 1482 |
| 8 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 1432 |
| 9 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 1276 |
| 10 | [Claude Opus 4.7 (Adaptive)](/models/claude-opus-4-7-adaptive) | Anthropic | 1241 |
| 11 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 1198 |

## FAQ

### What does Clocktower Bench measure?

How well a model plays Blood on the Clocktower, a social deduction game where a hidden evil team lies to the town. Playing good tests detection: weighing claims and catching liars. Playing evil tests deception: keeping a cover story straight and coordinating with teammates. The overall rating combines both.

### Which model leads Clocktower Bench?

In the 2026 study GPT-6 Astra leads at 1859, ahead of GPT-6.1 Sol at 1734. Claude Fable 5.1 is the top Anthropic model at 1595. OpenAI's lead comes mostly from detection: its town players keep updating on evidence across days, while Claude's towns tend to settle early.

### Why are the ratings not percentages?

The score is a relative rating fitted from head-to-head games, with 1500 set to the average of the 11 rated models. It only compares models inside this pool, so a new model's arrival can shift every number.

## Compare Top Models on Clocktower Bench

- [GPT-6 Astra vs GPT-6.1 Sol](/compare/gpt-6-1-sol-vs-gpt-6-astra)
- [GPT-6.1 Sol vs Claude Fable 5.1](/compare/claude-fable-5-1-vs-gpt-6-1-sol)
- [Claude Fable 5.1 vs GPT-5.6 Sol](/compare/claude-fable-5-1-vs-gpt-5-6-sol)
- [GPT-5.6 Sol vs Claude Opus 5.5](/compare/claude-opus-5-5-vs-gpt-5-6-sol)