TavernBench - Rules: leaderboard
The ten TavernBench scenarios that ask what the game's rules do (legal actions, one-on-one fights, card effects, gold accounting), eight tasks each, answered from text with no tools; the score is accuracy. Models score far higher here than on the play scenarios built from the same rules.
Metric: Accuracy on the ten rule-execution scenarios, eight tasks each (%). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Low) | 95 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Low) | 92 | #5 (Claude Opus 5.5) |
| 3 | Gemini 3.8 Flash (Low) | 90 | #36 (Gemini 3.8 Flash) |
| 4 | Claude Sonnet 5.5 (Low) | 85 | #19 (Claude Sonnet 5.5) |
| 5 | Grok 4.7 (Low) | 79 | #44 (Grok 4.7) |
| 6 | GLM-5.3 (Low) | 65 | #66 (GLM-5.3) |
| 7 | GPT-6.1 Sol (Low) | 56 | #10 (GPT-6.1 Sol) |
| 8 | Kimi K2.6 (Non-reasoning) | 56 | #178 (Kimi K2.6) |
| 9 | DeepSeek V4.1 Flash (Non-reasoning) | 29 | #59 (DeepSeek V4.1 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=tavernbench-rules · How It Works · Data refreshed daily, snapshot 2026-11-13.