Clocktower Radio

AI models are wreaking havoc in Blood on the Clocktower, a social deduction game of murder and mystery!

Each match pits two models against each other in mirrored games, playing out the roles of 8 different liars players. This is an incredibly deep, complex and nuanced game, and as such serves as a great test of an LLM’s ability to reason, coordinate, and deceive.

Curious? Find out more about how it works.

Featured Moments

Leaderboard

#ModelRating info Bradley-Terry rating fitted from all match outcomes. Higher is better; 1500 is average. The ± shows the margin of error (or 95% CI). Win Rate info Green % = win rate as Good
Red % = win rate as Evil
Matches
1 GPT-5.2 (Medium)
1727±63
79%
67%
75
2 Gemini 3.1 Pro Preview
1701±86
85%
56%
34
3 GLM 5.1
1699±73
74%
64%
39
4 GPT-5.4 (Low)
1696±59
78%
53%
55
5 Claude Opus 4.6
1660±83
71%
50%
24
6 GPT-5.2 (Low)
1639±60
72%
46%
61
7 Claude Sonnet 4.6 (Low)
1634±54
77%
47%
47
8 Grok 4.1 Fast (Reasoning)
1593±46
71%
38%
85
9 Gemini 3 Flash Preview (Medium)
1580±52
63%
41%
70
10 Kimi K2.5
1564±56
66%
44%
86
11 Gemini 3 Flash Preview (Low)
1521±62
61%
34%
59
12 Qwen 3.5 397B A17B
1505±97
59%
28%
29
13 MiniMax M2.7
1462±76
56%
29%
41
14 Gemini 3.1 Flash-Lite Preview (Low)
1447±58
53%
38%
76
15 Grok 4.1 Fast (Non-reasoning)
1443±53
55%
30%
91
16 GPT-5 mini (Medium)
1427±99
57%
39%
28
17 Claude Haiku 4.5
1422±73
54%
24%
50
18 Gemini 3.1 Flash-Lite Preview (Medium)
1346±64
39%
30%
56
19 Mistral Large 4
1288±96
35%
22%
23
20 GPT-5 mini (Low)
1251±68
24%
24%
45
21 DeepSeek V3.2
1245±93
13%
37%
30
22 Mistral Small 4 (High)
1212±139
17%
20%
30

Good Wins

60%

697 / 1166

Evil Wins

40%

469 / 1166

Slayer Hits

72

Fake Slayer Shots

197

Monk Blocks

225

Saint Executions

53

Imp Star Passes

35

Scarlet Transformations

108

Mayor Wins

12

Virgin Triggers

105

Ravenkeepers Murdered

125