Benchmarks
347 sources · 1,258 models × 3,038 benchmarks · 158,912 scoresValen Labs' FrontierBioML gives an agent 25 machine-learning tasks in computational biology, across antibody development, protein engineering, structural biology, binder design and functional genomics, such as selecting designed binders for lab testing or forecasting gene dependencies in unseen cancer types. The agent works alone in an offline sandbox (4 to 32 CPUs, no GPU, about six hours) and submits a prediction script, which is run twice in a fresh sandbox on held-out data such as new antibody families, targets or experiment rounds; a submission that passes every check is scored from the task's baseline (0) to a perfect answer (100). Each agent runs each task three times, runs blocked by a provider's safety filter or failing a check count as 0, and the score is the mean of the 25 task means. Scores are not comparable across tasks.
No signal or difficulty estimate. Too few models have been measured on this benchmark to estimate its difficulty.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high · OpenCode | 39.0 | |
| 2 | Gemini 4 Argon | high · OpenCode | 36.8 | |
| 3 | GPT-6.1 Sol | max · Codex | 29.1 | |
| 4 | GPT-6 Astra | max · Codex | 28.8 | |
| 5 | Muse Wonder | max · OpenCode | 26.9 | |
| 6 | Kimi K3.1 Pro | max · OpenCode | 25.8 | |
| 7 | Grok 4.7 | x-high · OpenCode | 24.0 | |
| 8 | Nemotron 3.5 Super VL 120B A12B (precision unspecified) | thinking · OpenCode | 11.0 |