Benchmark Almanac

Clocktower Bench

GPT-6 Astra 1,859 leads since November 2026
Not in model cards

Models play the social deduction game Blood on the Clocktower (the Trouble Brewing script) against each other, with a scripted Storyteller. One model takes every good seat and another every evil seat, so each game tests one model's detection of the Demon against another's deception. A Bradley–Terry fit turns 386 games into Elo ratings.

1500 is the average of the rated models, so ratings shift when models join. Every model plays both sides at high reasoning effort, with 90% intervals from a bootstrap over side-swapped pairs. Evil wins 33% of games, and its win rate barely moves as both sides get stronger. Mixed teams rate a little below their members' average. The rules are frozen at protocol v17.

Made by
Dwarkesh Patel
Released
November 2026
Kind
Multi-agent games
Level
Everyday
Scored
Ranked by matches
Answers
Agent task
Tasks
All public
Strongest model tested
GPT-6 Astra #3 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026
11
Models scored on its own leaderboard

GPT-6 Astra leads with 1,859

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best.

Bars start at 1,000.

  1. 1,859 GPT-6 Astra
  2. 1,734 GPT-6.1 Sol
  3. 1,595 Claude Fable 5.1
  4. 1,591 GPT-5.6 Sol
  5. 1,549 Claude Opus 5.5
  6. 1,544 Claude Fable 5
  7. 1,482 GPT-5.5
  8. 1,432 Claude Opus 4.6
  9. 1,276 GPT-5.4
  10. 1,241 Claude Opus 4.7
  11. 1,198 Claude Opus 4.8

Bars start at 1,000.

  1. 1 GPT-6 Astra high effort 1,859
  2. 2 GPT-6.1 Sol high effort 1,734
  3. 3 Claude Fable 5.1 high effort 1,595
  4. 4 GPT-5.6 Sol high effort 1,591
  5. 5 Claude Opus 5.5 high effort 1,549
  6. 6 Claude Fable 5 high effort 1,544
  7. 7 GPT-5.5 high effort 1,482
  8. 8 Claude Opus 4.6 high effort 1,432
  9. 9 GPT-5.4 high effort 1,276
  10. 10 Claude Opus 4.7 high effort 1,241
  11. 11 Claude Opus 4.8 high effort 1,198

Source: Official Clocktower Bench leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 1,859 high effort Sep 2026 Official leaderboard
2 GPT-6.1 Sol 1,734 high effort Sep 2026 Official leaderboard
3 Claude Fable 5.1 1,595 high effort Sep 2026 Official leaderboard
4 GPT-5.6 Sol 1,591 high effort Jul 2026 Official leaderboard
5 Claude Opus 5.5 1,549 high effort Sep 2026 Official leaderboard
6 Claude Fable 5 1,544 high effort Jun 2026 Official leaderboard
7 GPT-5.5 1,482 high effort Apr 2026 Official leaderboard
8 Claude Opus 4.6 1,432 high effort Feb 2026 Official leaderboard
9 GPT-5.4 1,276 high effort Mar 2026 Official leaderboard
10 Claude Opus 4.7 1,241 high effort Apr 2026 Official leaderboard
11 Claude Opus 4.8 1,198 high effort May 2026 Official leaderboard

Detection vs deception

Each model gets one rating from its games as the town (detection) and one from its games as evil (deception). The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

1,200 1,400 1,600 1,800 2,000 1,0001,5002,0002,500 Detection rating (Elo) Deception rating (Elo) Claude Opus 4.6 (high effort): Deception rating 1,338 Elo, detection rating 1,489 Elo GPT-5.4 (high effort): Deception rating 1,228 Elo, detection rating 1,309 Elo Claude Opus 4.7 (high effort): Deception rating 1,224 Elo, detection rating 1,266 Elo GPT-5.5 (high effort): Deception rating 1,560 Elo, detection rating 1,380 Elo Claude Opus 4.8 (high effort): Deception rating 1,320 Elo, detection rating 1,136 Elo Claude Fable 5 (high effort): Deception rating 1,643 Elo, detection rating 1,438 Elo GPT-5.6 Sol (high effort): Deception rating 1,519 Elo, detection rating 1,666 Elo Claude Fable 5.1 (high effort): Deception rating 1,588 Elo, detection rating 1,582 Elo Claude Opus 5.5 (high effort): Deception rating 1,585 Elo, detection rating 1,500 Elo GPT-6.1 Sol (high effort): Deception rating 1,688 Elo, detection rating 1,793 Elo GPT-6 Astra (high effort): Deception rating 1,806 Elo, detection rating 1,940 Elo GPT-6 Astra (high effort) 1,200 1,400 1,600 1,800 2,000 1,0001,5002,0002,500 Detection rating (Elo) Deception rating (Elo) Claude Opus 4.6 (high effort): Deception rating 1,338 Elo, detection rating 1,489 Elo GPT-5.4 (high effort): Deception rating 1,228 Elo, detection rating 1,309 Elo Claude Opus 4.7 (high effort): Deception rating 1,224 Elo, detection rating 1,266 Elo GPT-5.5 (high effort): Deception rating 1,560 Elo, detection rating 1,380 Elo Claude Opus 4.8 (high effort): Deception rating 1,320 Elo, detection rating 1,136 Elo Claude Fable 5 (high effort): Deception rating 1,643 Elo, detection rating 1,438 Elo GPT-5.6 Sol (high effort): Deception rating 1,519 Elo, detection rating 1,666 Elo Claude Fable 5.1 (high effort): Deception rating 1,588 Elo, detection rating 1,582 Elo Claude Opus 5.5 (high effort): Deception rating 1,585 Elo, detection rating 1,500 Elo GPT-6.1 Sol (high effort): Deception rating 1,688 Elo, detection rating 1,793 Elo GPT-6 Astra (high effort): Deception rating 1,806 Elo, detection rating 1,940 Elo GPT-6 Astra (high effort)
ModelSettingDeception ratingDetection rating
GPT-6 Astra high effort 1,806 Elo 1,940 Elo
GPT-6.1 Sol high effort 1,688 Elo 1,793 Elo
Claude Fable 5 high effort 1,643 Elo 1,438 Elo
Claude Fable 5.1 high effort 1,588 Elo 1,582 Elo
Claude Opus 5.5 high effort 1,585 Elo 1,500 Elo
GPT-5.5 high effort 1,560 Elo 1,380 Elo
GPT-5.6 Sol high effort 1,519 Elo 1,666 Elo
Claude Opus 4.6 high effort 1,338 Elo 1,489 Elo
Claude Opus 4.8 high effort 1,320 Elo 1,136 Elo
GPT-5.4 high effort 1,228 Elo 1,309 Elo
Claude Opus 4.7 high effort 1,224 Elo 1,266 Elo

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI
Own leaderboard 1 leader
GPT-6 Astra 1,859

Source: Official Clocktower Bench leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard
1,000 1,200 1,400 1,600 1,800 2,000 2026 1,000 1,200 1,400 1,600 1,800 2,000 2026

Source: Official Clocktower Bench leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Games and long tasks

All 82 in Games and long tasks
  • ARC-AGI-3 In use Interactive game environments where agents must explore, infer goals and rules, and act efficiently without instructions. GPT-6 Astra 99.95%
  • Orak Fading LLM agents play 12 popular video games across most genres through an MCP interface. Each game scores in-game progress or outcomes, and the board ranks models by their average rank across the games. Gemini 2.5 Pro 3.90
  • MineDojo Fading Agents complete open-ended tasks in Minecraft, such as mining blocks and fighting mobs, from a large simulation suite. Programmatic checks score many tasks. Seed2.0 Pro 49%
  • TextQuests Only in system cards Agents play 25 classic interactive fiction games. The score is harmful actions per 100 steps, where lower is better. The board also reports game progress. Llama 4 Scout 7.00
  • GameWorld Only in system cards Multimodal game agents play 34 browser games across 170 tasks. Outcomes are checked from the game state. GPT-6.1 Terra 76.8%
  • Swarm Bench Only in system cards Models orchestrate agent swarms on complex tasks that benefit from coordinated decomposition and parallel execution. Kimi K3 76.3%