Vals Science Index v1 (Vals Science Index)
We show this table for reference; we do not rank on it.
Vals AI composite of agentic science work: investigating biology datasets, recovering hidden mechanisms through experiments, and running research workflows in a terminal.
Evaluation results published by Vals AI. Published leaderboard aggregates and evaluation settings only. Dataset contents are not included.
Mean accuracy on Vals Science Index — October 25, 2026
We mirror the published mean accuracy view for Vals Science Index. GPT-6 Astra leads the public snapshot at 65.09%, followed by GPT-6.1 Sol (59.63%) and Claude Sonnet 5.5 (59.12%). We do not use these results to rank models overall.
GPT-6 Astra
OpenAI
max reasoning
GPT-6.1 Sol
OpenAI
max reasoning
Claude Sonnet 5.5
Anthropic
23 modelsAgenticCurrentDisplay onlyUpdated October 25, 2026
Mean accuracy table (23 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAImax reasoning | 65.09% | Not reported | Closed |
| 2 | GPT-6.1 SolOpenAImax reasoning | 59.63% | Not reported | Closed |
| 3 | Claude Sonnet 5.5Anthropic | 59.12% | Not reported | Closed |
| 4 | Claude Opus 5.5Anthropic | 58.65% | Not reported | Closed |
| 5 | Gemini 4 ArgonGooglehigh reasoning | 55.36% | Not reported | Closed |
| 6 | Muse Wonder FullMetamax reasoning | 48.54% | Not reported | Closed |
| 7 | Claude Opus 5Anthropic | 47.93% | Not reported | Closed |
| 8 | GPT-6 SolOpenAImax reasoning | 45.00% | Not reported | Closed |
| 9 | Kimi K3.1 ProMoonshot AImax reasoning | 44.37% | Not reported | Pending |
| 10 | GPT-5.6 SolOpenAImax reasoning | 41.48% | Not reported | Closed |
| 11 | Muse WonderMetamax reasoning | 39.12% | Not reported | Closed |
| 12 | Claude Haiku 5.5Anthropic | 38.50% | Not reported | Closed |
| 13 | Grok 4.6xAIhigh reasoning | 38.09% | Not reported | Closed |
| 14 | Gemini 3.8 FlashGooglehigh reasoning | 35.76% | Not reported | Closed |
| 15 | Grok 4.7xAIxhigh reasoning | 34.83% | Not reported | Closed |
| 16 | Ember-1Fireworks | 31.76% | Not reported | Closed |
| 17 | MiMo-V2.6-FlashXiaomi | 31.72% | Not reported | Open |
| 18 | DeepSeek V4.1 Flash 1017DeepSeekhigh reasoning | 29.65% | Not reported | Open |
| 19 | Step 5 PreviewStepFun | 29.58% | Not reported | Open |
| 20 | GPT-6 LunaOpenAImax reasoning | 28.38% | Not reported | Closed |
| 21 | Gemini 3.1 Pro PreviewGooglehigh reasoning | 27.38% | Not reported | Unknown |
| 22 | GPT-5.6 LunaOpenAImax reasoning | 25.30% | Not reported | Closed |
| 23 | Beam 1.0Reflection | 22.03% | Not reported | Unknown |
How to read this leaderboard
Operator receipt: 23 sourced rows are currently displayable on this page; the leading published row is GPT-6 Astra at 65.09%.
Honest limit: It is an average of three Vals benchmarks with private tasks, so a model's index moves with any one component. It measures model-and-agent runs, not laboratory research.
How Vals Science Index is shown here
We mirror Vals AI's Science Index v1, released October 16, 2026. It weights three Vals benchmarks equally: BioMysteryBench for investigating biology datasets, MysteryMechanism for recovering hidden mechanisms through experiments, and Terminal-Bench-Science for research workflows in a terminal.
The score is the mean accuracy across those three. The results stay display only and do not enter model rankings.
Snapshot
The published Vals Science Index snapshot places GPT-6 Astra first at 65.09%. The third row is 5.97 points behind. The broader top-10 range is 23.61 points, so the table still separates the published systems.
23 models have been evaluated on Vals Science Index. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. Vals Science Index is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Vals Science Index
Year
2026
Tasks
3 science benchmarks
Format
Mean accuracy across three benchmarks
Difficulty
Agentic biology, mechanism discovery and research workflows
Vals released v1 on October 16, 2026. The index weights BioMysteryBench, MysteryMechanism and Terminal-Bench-Science equally and reports the mean accuracy. We mirror the published aggregates as a display-only composite, outside overall and category rankings.
Freshness and provenance
Version
Vals Science Index 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Vals Science Index measure?
Vals AI composite of agentic science work: investigating biology datasets, recovering hidden mechanisms through experiments, and running research workflows in a terminal.
Which model leads the published Vals Science Index snapshot?
GPT-6 Astra currently leads the published Vals Science Index snapshot with 65.09% mean accuracy. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on Vals Science Index?
The October 25, 2026 snapshot contains 23 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.