Deep20Bench: leaderboard
Play Twenty Questions: identify a hidden subject by asking informative questions and using the oracle answers to narrow possible subjects. Deep20Bench edition 1.0: 7 fixed public subjects, 5 rounds per subject, 50-question limit. Question score averages counted turns within each subject, then equally across subjects; model failures score 51. Lower is better. The primary selects the latest complete accepted run, never the best score, including qualified experimental runs. Output-format penalties remain in native scores. Edition-specific answer rules and accepted prompt revisions are retained in source provenance; scores across editions are not directly comparable.
Metric: Question score: subject-mean counted questions including model-failure penalty (lower is better). Source: mindalyze-com.github.io. Saturation forecast: Around 2029. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (High) | 12.06 | #20 (Claude Fable 5) |
| 2 | Claude Fable 5.1 (High) | 12.11 | #10 (Claude Fable 5.1) |
| 3 | Claude Opus 5 (High) | 12.34 | #16 (Claude Opus 5) |
| 4 | Kimi K3 (High) | 12.74 | #35 (Kimi K3) |
| 5 | Gemini 3.8 Flash (High) | 13.14 | #30 (Gemini 3.8 Flash) |
| 6 | GPT-OSS-120B (High) | 13.63 | #346 (GPT-OSS-120B) |
| 7 | Gemini 3.7 Flash (High) | 14.03 | #24 (Gemini 3.7 Flash) |
| 8 | Grok 4.6 (High) | 14.29 | #29 (Grok 4.6) |
| 9 | GPT-5.6 Sol (High) | 14.4 | #22 (GPT-5.6 Sol) |
| 10 | Claude Sonnet 5 (High) | 14.74 | #65 (Claude Sonnet 5) |
| 11 | GPT-6 Astra (High) | 15.09 | #6 (GPT-6 Astra) |
| 12 | Grok 4.5 (High) | 15.17 | #54 (Grok 4.5) |
| 13 | GPT-5 Nano (Medium) | 15.49 | #439 (GPT-5 Nano) |
| 14 | Gemini 3.6 Flash (High) | 16.89 | #57 (Gemini 3.6 Flash) |
| 15 | GLM-5.3 Flash (High) | 17.6 | #77 (GLM-5.3 Flash) |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=deep20bench · How It Works · Data refreshed daily, snapshot 2026-11-05.