Clocktower Radio
Models play the social deduction game Blood on the Clocktower (the Trouble Brewing script) in mirrored pairs of games. In each match one model plays good and the other evil, and then they swap. A coded Storyteller runs the game, and a Bradley–Terry fit over all matches gives ratings where 1500 is average.
The board also reports each model's win rate as good and as evil, its cost per game and its output tokens per action. Good wins 60% of games. The changelog shows rule fixes through July 27th, 2026. The release month is the first changelog entry.
- Made by
- Che Jami
- Released
- March 2026
- Kind
- Multi-agent games
- Level
- Everyday
- Scored
- Ranked by matches
- Answers
- Agent task
- Tasks
- All public
- Strongest model tested
- GPT-5.5 #20 on the capability index
- Last model added
- October 2026
- Scores refreshed
- November 12th, 2026
GPT-5.5 leads with 1,708
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. Every run is in the next chart.
- 1,708 GPT-5.5
- 1,706
Gemini 3.1 Pro - 1,698 Gemini 3.5 Flash
- 1,696 Kimi K2.6
- 1,694 GPT-5.2
- 1,684 MiMo-V2.5-Pro
- 1,661 GPT-5.4
- 1,661
DeepSeek-V4-Pro - 1,655 GLM-5.1
- 1,625 Claude Opus 4.6
- 1,615 Claude Sonnet 4.6
- 1,552 Grok 4.1 Fast
- 1,541 Gemini 3 Flash
- 1,530 Grok 4.3
- 1,517 Kimi K2.5
- 1,478
Qwen3.5 397B-A17B - 1,415
MiniMax-M2.7 - 1,405 Gemini 3.1 Flash-Lite
- 1,393 GPT-5 mini
- 1,390 Claude Haiku 4.5
- 1,265
Mistral Large 4 - 1,242 Mistral Small 4
- 1,213 DeepSeek-V3.2
Bars start at 1,200.
- 1 GPT-5.5 1,708
- 2 Gemini 3.1 Pro 1,706
- 3 Gemini 3.5 Flash 1,698
- 4 Kimi K2.6 1,696
- 5 GPT-5.2 medium 1,694
- 6 MiMo-V2.5-Pro 1,684
- 7 GPT-5.4 low 1,661
- 8 DeepSeek-V4-Pro 1,661
- 9 GLM-5.1 1,655
- 10 Claude Opus 4.6 1,625
- 11 Claude Sonnet 4.6 low 1,615
- 12 Grok 4.1 Fast reasoning 1,552
- 13 Gemini 3 Flash medium 1,541
- 14 Grok 4.3 1,530
- 15 Kimi K2.5 1,517
- 16 Qwen3.5 397B-A17B 1,478
- 17 MiniMax-M2.7 1,415
- 18 Gemini 3.1 Flash-Lite low 1,405
- 19 GPT-5 mini medium 1,393
- 20 Claude Haiku 4.5 1,390
- 21 Mistral Large 4 1,265
- 22 Mistral Small 4 high 1,242
- 23 DeepSeek-V3.2 1,213
Source: Official Clocktower Radio leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 1,708 | Apr 2026 | Official leaderboard | |
| 2 | Gemini 3.1 Pro | 1,706 | Feb 2026 | Official leaderboard | |
| 3 | Gemini 3.5 Flash | 1,698 | May 2026 | Official leaderboard | |
| 4 | Kimi K2.6 | 1,696 | Apr 2026 | Official leaderboard | |
| 5 | GPT-5.2 | 1,694 | medium | Dec 2025 | Official leaderboard |
| 6 | MiMo-V2.5-Pro | 1,684 | Apr 2026 | Official leaderboard | |
| 7 | GPT-5.4 | 1,661 | low | Mar 2026 | Official leaderboard |
| 8 | DeepSeek-V4-Pro | 1,661 | Apr 2026 | Official leaderboard | |
| 9 | GLM-5.1 | 1,655 | Apr 2026 | Official leaderboard | |
| 10 | Claude Opus 4.6 | 1,625 | Feb 2026 | Official leaderboard | |
| 11 | Claude Sonnet 4.6 | 1,615 | low | Feb 2026 | Official leaderboard |
| 12 | Grok 4.1 Fast | 1,552 | reasoning | Nov 2025 | Official leaderboard |
| 13 | Gemini 3 Flash | 1,541 | medium | Dec 2025 | Official leaderboard |
| 14 | Grok 4.3 | 1,530 | Apr 2026 | Official leaderboard | |
| 15 | Kimi K2.5 | 1,517 | Jan 2026 | Official leaderboard | |
| 16 | Qwen3.5 397B-A17B | 1,478 | Feb 2026 | Official leaderboard | |
| 17 | MiniMax-M2.7 | 1,415 | Mar 2026 | Official leaderboard | |
| 18 | Gemini 3.1 Flash-Lite | 1,405 | low | Mar 2026 | Official leaderboard |
| 19 | GPT-5 mini | 1,393 | medium | Aug 2025 | Official leaderboard |
| 20 | Claude Haiku 4.5 | 1,390 | Oct 2025 | Official leaderboard | |
| 21 | Mistral Large 4 | 1,265 | Oct 2026 | Official leaderboard | |
| 22 | Mistral Small 4 | 1,242 | high | Mar 2026 | Official leaderboard |
| 23 | DeepSeek-V3.2 | 1,213 | Dec 2025 | Official leaderboard |
The best score, 1,708, costs $3.45 a task
Runs on the cost frontier score higher than any run that's cheaper. DeepSeek-V4-Pro comes within 47 of the best for $0.31 a task, 11 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official Clocktower Radio leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.5 | 1,708 | $3.45 | |||
| Gemini 3.1 Pro | 1,706 | $3.15 | |||
| Gemini 3.5 Flash | 1,698 | $2.26 | |||
| Kimi K2.6 | 1,696 | $2.95 | |||
| GPT-5.2 | medium | 1,694 | $1.97 | ||
| MiMo-V2.5-Pro | 1,684 | $0.36 | |||
| GPT-5.4 | low | 1,661 | $2.51 | ||
| DeepSeek-V4-Pro | 1,661 | $0.31 | |||
| GLM-5.1 | 1,655 | $0.94 | |||
| Claude Opus 4.6 | 1,625 | $3.71 | |||
| Claude Sonnet 4.6 | low | 1,615 | $1.82 | ||
| GPT-5.2 | low | 1,612 | $1.27 | ||
| Grok 4.1 Fast | reasoning | 1,552 | $0.19 | ||
| Gemini 3 Flash | medium | 1,541 | $0.34 | ||
| Grok 4.3 | 1,530 | $1.43 | |||
| Kimi K2.5 | 1,517 | $0.39 | |||
| Gemini 3 Flash | low | 1,478 | $0.28 | ||
| Qwen3.5 397B-A17B | 1,478 | $1.06 | |||
| Grok 4.1 Fast | non-reasoning | 1,415 | $0.10 | ||
| MiniMax-M2.7 | 1,415 | $0.27 | |||
| Gemini 3.1 Flash-Lite | low | 1,405 | $0.17 | ||
| GPT-5 mini | medium | 1,393 | $0.53 | ||
| Claude Haiku 4.5 | 1,390 | $0.71 | |||
| Gemini 3.1 Flash-Lite | medium | 1,313 | $0.16 | ||
| Mistral Large 4 | 1,265 | $0.27 | |||
| Mistral Small 4 | high | 1,242 | $0.12 | ||
| DeepSeek-V3.2 | 1,213 | $0.32 | |||
| GPT-5 mini | low | 1,206 | $0.24 |
2 models have held number one on the official leaderboard since its release in March 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official Clocktower Radio leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Clocktower Radio 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
- 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%