BenchBench-Cloning
Science · 2026-10-23
65 molecular-cloning planning tasks from Benchling and DNAForge grounded in sequence records, lab notebooks and published experiments: 50 protocol-development tasks (pass if every deterministic check passes; ten attempts each) and 15 construct-design tasks (pass if the final score is at least 0.9; five attempts each). Agents run in their native harnesses in sandboxes with network access limited to the model provider. The score is the task-weighted aggregate pass rate (50/65 protocol, 15/65 construct), in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 30.6 |
| GPT-6.1 Sol | 22.2 |
| Claude Opus 5.5 | 20.5 |
| Gemini 4 Argon | 18.8 |
| Kimi K3 | 11.1 |
| Grok 4.7 | 10.9 |
| Gemini 3.8 Flash | 10.8 |