BenchBench-Protocol (Best of 10): leaderboard
BenchBench-Protocol (Benchling, 2026) derives 149 wet-lab protocol-modification tasks from changes scientists made to published protocols, each graded against weighted rubric elements by a GPT-5.6 Terra judge. This board gives the expected best-of-10 score: for each task the highest rubric score among ten sampled responses, averaged over tasks, in percent; higher is better.
Metric: Expected best-of-10 normalized rubric score (%; per task the best of ten sampled responses by weighted rubric score, 149 protocol-modification tasks, graded by GPT-5.6 Terra). Source: arxiv.org. Saturation forecast: Around March 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Max) | 75.3 | #20 (Claude Opus 5) |
| 2 | Kimi K3 (Max) | 62.8 | #53 (Kimi K3) |
| 3 | GPT-5.6 Terra (Max) | 57.7 | #60 (GPT-5.6 Terra) |
| 4 | GPT-5.6 Sol (Max) | 57.5 | #29 (GPT-5.6 Sol) |
| 5 | GLM-5.2 (Max) | 54.9 | #137 (GLM-5.2) |
| 6 | Inkling (Max) | 54.1 | #270 (Inkling) |
| 7 | Grok 4.5 (High) | 53.3 | #78 (Grok 4.5) |
| 8 | GPT-5.6 Luna (Max) | 52.7 | #117 (GPT-5.6 Luna) |
| 9 | Gemini 3.6 Flash (High) | 47.6 | #72 (Gemini 3.6 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=benchbench-protocol-best-of-10 · How It Works · Data refreshed daily, snapshot 2026-11-12.