BioMysteryBench (Vals Terminus 2)
Science · 2026-09-22
Ninety bioinformatics mysteries evaluated by Vals with Terminus 2. Three trials per task; solved proportion averaged across trials. Content-filter, refusal and protocol failures count as failures.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 79.3 |
| Claude Opus 5 | 79.3 |
| GPT-6 Astra | 79.3 |
| GPT-6 Sol | 74.8 |
| Grok 4.6 | 72.2 |
| Kimi K3 | 71.5 |
| GPT-5.6 Sol | 71.1 |
| Grok 4.7 | 69.3 |
| Hy4 Preview | 69.3 |
| DeepSeek V4.1 Flash | 67.8 |
| Muse Spark 1.2 | 64.8 |
| DeepSeek-V4-Flash-0731 | 64.4 |
| Gemini 3.8 Flash | 62.2 |
| GPT-5.6 Luna | 61.5 |
| GPT-6 Luna | 61.5 |
| Gemini 3.6 Flash | 58.5 |