DrBencher: leaderboard
DrBencher (2026) generates questions that need multi-hop entity identification, property retrieval from authoritative sources and domain computation, built answer-first from knowledge-graph chains, and evaluates models closed-book on a human-validated set. This board is answer accuracy in percent over all human-validated questions, the mean of three runs, with a 2 percent relative tolerance for numbers; higher is better.
Metric: Answer accuracy (%) on the 268 human-validated DrBencher questions (closed-book, default API temperature, mean of three runs, 2% relative tolerance and exact match for history); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 4.6 | 20.1 | #113 | |
| 2 | Gemini 2.5 Flash | 18.4 | #463 | |
| 3 | Llama 4 Maverick | 17.9 | #1110 | |
| 4 | GPT-5.2 | 10.6 | #185 | |
| 5 | Qwen 3 30B A3B | Thinking | 10 | #1156 |
| 6 | Mistral Small 3.2 | 6.8 | #1641 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=drbencher · How It Works · Data refreshed daily, snapshot 2026-11-21.