Benchmarks
347 sources · 1,258 models × 3,038 benchmarks · 158,912 scoresTests whether agents can reason correctly from the 3D structures of pathogen molecules that guide vaccine, antibody and antiviral design, such as identifying the metal ions in a viral enzyme, finding which parts of a viral surface protein an antibody can reach, or confirming that a candidate drug is really bound. Each of the 58 evaluations pairs crystal, cryo-EM, predicted-model or assay data with a question and an expert-validated answer; scores are the mean pass rate over three runs per evaluation, with refusals and errors counted as failures. benchmarks.bio excludes it from its overall Capabilities score.
601st strongest of 798 fitted benchmarks
Signal-to-noise +3.50
15th hardest of 798 fitted benchmarks
Capability needed ≳ 193.5
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | not reported · Claude Code | 43.7 | |
| 2 | Claude Opus 5.5 | not reported · Claude Code | 31.0 | |
| 3 | DeepSeek V4.1 Flash | not reported · pi | 27.6 | |
| 4 | GPT-6.1 Terra | not reported · pi | 25.9 | |
| 5 | Mistral Large 4 | not reported · pi | 23.0 | |
| 6 | Claude Sonnet 5.5 | not reported · Claude Code | 21.8 | |
| 7 | Kimi K3 | not reported · pi | 21.8 | |
| 8 | GPT-6.1 Sol | not reported · pi | 21.3 | |
| 9 | GPT-6 Astra | not reported · Codex | 21.3 | |
| 10 | Claude Opus 5 | not reported · Claude Code | 20.7 | |
| 11 | Sonnet 5 | not reported · Claude Code | 20.7 | |
| 12 | GPT-5.6 Sol | not reported · pi | 20.7 | |
| 13 | GPT-5.6 Luna | not reported · pi | 20.1 | |
| 14 | GLM-5.3 | not reported · pi | 19.0 | |
| 15 | Grok 4.6 | not reported · pi | 15.5 | |
| 16 | Gemini 4 Argon | not reported · pi | 14.9 | |
| 17 | GPT-6 Sol | not reported · pi | 13.8 | |
| 18 | GPT-6 Luna | not reported · pi | 13.2 | |
| 19 | Claude Fable 5.1 | not reported · pi | 6.3 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).