Clocktower Radio: leaderboard

Models play the social-deduction game Blood on the Clocktower (Trouble Brewing script, 22 roles) against each other with an automated rules engine as storyteller; each match is two games between two models, the second a mirror with roles kept and teams flipped, and the board is a Bradley-Terry rating over all match outcomes (1500 = average). A separate site by Che Jami, not the Clocktower Bench study; the table keeps configurations the site has superseded, rated in the same fit.

Metric: Bradley-Terry rating (1500 = average). Source: clocktower-radio.com. Saturation forecast: Around July 2027. 22 models tracked. Updated 19 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.51708#54
2Gemini 3.1 Pro (Preview)1706#101
3Gemini 3.5 Flash1698#108
4Kimi K2.61696#192
5GPT-5.2Medium1694#187
6MiMo-V2.5-Pro1684#289
7DeepSeek V4 Pro1661#198
8GPT-5.4Low1661#138
9GLM-5.11655#197
10Claude Opus 4.61625#114
11Claude Sonnet 4.6Low1615#175
12Grok 4.1 FastReasoning1552#433
13Gemini 3 Flash (Preview)Medium1541#156
14Grok 4.31530#338
15Kimi K2.51517#265

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=clocktower-radio · How It Works · Data refreshed daily, snapshot 2026-11-20.