TavernBench: leaderboard
Language models answer Hearthstone Battlegrounds decisions (combat outcomes, positioning, play order, targets, card choices, turn plans) from a text description of the game state, with no tools. Labels use simulator outcomes, and top human players' placement statistics for hero and trinket choices. The score is accuracy over the 30 play scenarios of a 368-task set, which asks for the best play rather than what the rules do.
Metric: Play accuracy: share of tasks answered correctly, mean over the 30 play scenarios (%). Source: arxiv.org. Saturation forecast: Around January 2028. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Low) | 63 | #6 (GPT-6 Astra) |
| 2 | Claude Sonnet 5.5 (Low) | 53 | #19 (Claude Sonnet 5.5) |
| 3 | Claude Opus 5.5 (Low) | 52 | #5 (Claude Opus 5.5) |
| 4 | GPT-6.1 Sol (Low) | 46 | #10 (GPT-6.1 Sol) |
| 5 | Gemini 3.8 Flash (Low) | 44 | #36 (Gemini 3.8 Flash) |
| 6 | GLM-5.3 (Low) | 40 | #66 (GLM-5.3) |
| 7 | Kimi K2.6 (Non-reasoning) | 28 | #178 (Kimi K2.6) |
| 8 | DeepSeek V4.1 Flash (Non-reasoning) | 26 | #59 (DeepSeek V4.1 Flash) |
| 9 | Grok 4.7 (Low) | 22 | #44 (Grok 4.7) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=tavernbench · How It Works · Data refreshed daily, snapshot 2026-11-13.