Vals Science Index
Indexes · 2026-10-16
Vals Science Index v1 is the unweighted mean of three agentic science benchmarks run by Vals: BioMysteryBench (90 bioinformatics mysteries with Terminus 2, three trials per task), MysteryMechanism (222 sealed mechanisms recovered through bounded experiments) and Terminal-Bench-Science 0.1 (70 research workflows with the row-disclosed agent harness and infrastructure retries). Component runs use their respective harnesses and match the standalone Vals leaderboards, which Atlas imports separately; mixed-model fallback aggregates are excluded from single-model observations.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 65.1 |
| GPT-6.1 Sol | 59.6 |
| Gemini 4 Argon | 55.4 |
| Muse Wonder Full | 48.5 |
| Claude Opus 5 | 47.9 |
| GPT-6 Sol | 45.0 |
| GPT-5.6 Sol | 41.5 |
| Muse Wonder | 39.1 |
| Claude Haiku 5.5 | 38.5 |
| Grok 4.6 | 38.1 |
| Gemini 3.8 Flash | 35.8 |
| Grok 4.7 | 34.8 |
| Ember-1 | 31.8 |
| MiMo-V2.6-Flash | 31.7 |
| DeepSeek V4.1 Flash | 29.6 |
| Step 5 Preview | 29.6 |
| GPT-6 Luna | 28.4 |
| Gemini 3.1 Pro Preview | 27.4 |
| GPT-5.6 Luna | 25.3 |
| Beam 1.0 501B A23B | 22.0 |