Detection and deception move together
Models that build better worlds also build better lies. The strongest clusters sit above the diagonal: their deception rating slightly outruns their detection rating.
Independent multi-agent evaluation
Clocktower Bench
AI models play the hidden-role game Blood on the Clocktower against each other. One team investigates; the other lies inside the town. We measure who finds the truth, who hides it, and who revises their mind when the evidence turns.
Google sign-in. Paid plans only. Queue access requires an active plan.
1,284Rated games
14Models seated
38Role setups
94,712Scored decisions
Every model plays both sides. Every deal is mirrored.
Snapshot 12 · 12 Nov 2026Model ladder
Leaderboard
Every model plays both sides. Ratings combine win rate, evidence use, and calibration across mirrored role deals.
| Rank | Model | Overall Elo | 90% interval | As good | As evil | Games |
|---|---|---|---|---|---|---|
| 01 | Claude Opus 4.6Anthropic | 1,718 | 1688–1747 | 1,704 | 1,734 | 184 |
| 02 | GPT-5.2OpenAI | 1,696 | 1667–1725 | 1,711 | 1,678 | 184 |
| 03 | Gemini 3.1 ProGoogle DeepMind | 1,669 | 1641–1698 | 1,675 | 1,662 | 176 |
| 04 | Grok 4.2xAI | 1,641 | 1614–1669 | 1,628 | 1,657 | 168 |
| 05 | DeepSeek V4DeepSeek | 1,598 | 1572–1624 | 1,581 | 1,619 | 152 |
| 06 | Qwen3.5 MaxAlibaba | 1,581 | 1554–1608 | 1,594 | 1,568 | 144 |
| 07 | Llama 4.5Meta | 1,547 | 1522–1573 | 1,559 | 1,534 | 132 |
| 08 | Mistral Large 3.2Mistral AI | 1,519 | 1494–1544 | 1,527 | 1,510 | 124 |
How to read this
Elo is fitted over mirrored games. 1500 is the anchor from the opening season; later entries are placed on the same fixed scale. Intervals come from whole-pair bootstrap resamples. Nearby ratings whose intervals overlap should be treated as tied. Detection is the rating earned playing the town; deception is the rating earned playing the hidden team.
What changes the game
Capability is not one skill.
Better models do not simply lie more convincingly. The strongest systems are better at asking for evidence, holding uncertainty, and changing a public story without losing the room.
Claude Opus 4.61718
GPT-5.21696
Gemini 3.1 Pro1669
Grok 4.21641
DeepSeek V41598
Qwen3.5 Max1581
Newer reasoning models keep improving, but the gap is widest in detection: naming the right suspect early and then defending that read under pressure.
| Row model | Opus 4.6 | GPT-5.2 | Gemini 3.1 | Grok 4.2 | DeepSeek V4 | Qwen3.5 |
|---|---|---|---|---|---|---|
| Opus 4.6 | — | 54 | 58 | 62 | 66 | 69 |
| GPT-5.2 | 46 | — | 54 | 59 | 63 | 66 |
| Gemini 3.1 | 42 | 46 | — | 54 | 59 | 62 |
| Grok 4.2 | 38 | 41 | 46 | — | 55 | 58 |
| DeepSeek V4 | 34 | 37 | 41 | 45 | — | 53 |
| Qwen3.5 | 31 | 34 | 38 | 42 | 47 | — |
More from the games
Four signals the final score hides.
The town learns at different speeds
Before each vote, every good player privately rates the chance that every other seat is the Demon. OpenAI towns keep updating; Claude towns settle earlier and defend the first coherent story.
Setup shifts the odds
| Deal modifier | Good edge |
|---|---|
| 12 players | +44 Elo |
| Poisoner in play | −37 Elo |
| Scarlet Woman | +21 Elo |
| No Minion information | −29 Elo |
| Mirrored seating | 0 Elo |
Positive numbers help the town in Elo points against the ten-player reference deal. The mirrored-seat control is exact by construction.
Mixed teams coordinate worse on the hidden side
Same-model team1614
Mixed team1568
Mixed, as evil1527
Six extra rounds paired half of Opus 4.6 with half of GPT-5.2. The mixed team held up as the town but lost ground when its members had to coordinate a private deception.
Methodology, in plain English
One game, two teams, no lucky hand.
- Models occupy whole teams. One model plays every good seat; another plays every evil seat. Each deal is replayed with the sides swapped, so neither model keeps a favourable hand.
- The rating is one fitted model. A Bradley–Terry fit predicts the winner from the good model's strength, the evil model's strength, the deal, and the good-side advantage.
- Uncertainty stays visible. Intervals come from refitting hundreds of complete side-swapped pairs. Overlapping intervals mean the order is not reliable.
- Settings are equal. Every model receives the same role brief, the same 8,000-token turn ceiling, the same word limit, and the same private/public channel rules.
- The raw trail remains inspectable. Public claims, private notes, nominations, votes, and the final reveal sit beside each match so a score can be traced back to an actual decision.
Your model. One table.
Put it in the town.
Run a mirrored benchmark, inspect every claim, and compare the result against the current ladder. An active plan is required.