Proprietary

Vals Science Index

Updated 10/16/2026

A single measure of AI's ability to do science — agentic model performance on biology data investigation, mechanism discovery, and research workflows, weighted equally.

As of October 16, 2026, GPT-6 Astra ranks first on the Vals Science Index with 65.09%, followed by GPT-6.1 Sol (59.63%) and Claude Sonnet 5.5 (59.12%).

Vals Science IndexAgentic science benchmark
ACCURACY

Vals Science Index leaderboard

Rank Model Accuracy Cost / Test Input / Output Cost Input / Output Tokens Duration
1 GPT-6 Astra 65.09% $8.03 $10 / $50 884.86M / 25.81M 40m37s
2 GPT-6.1 Sol 59.63% $1.36 $2 / $10 757.08M / 27.12M 33m11s
3 Claude Sonnet 5.5 59.12% $13.24 $2 / $10 20.05B / 115.89M 1h09m
4 Claude Opus 5.5 58.65% $9.13 $4 / $20 4.68B / 64.76M 59m51s
5 Gemini 4 Argon 55.36% $18.40 $4 / $20 4.96B / 70.14M 57m00s
6 Wonder Full 48.54% $10.15 $4 / $20 5.13B / 45.14M 1h34m
7 Claude Opus 5 47.93% $13.04 $5 / $25 N/A 1h02m
8 GPT-6 Sol 45.00% $2.31 $2 / $10 2.42B / 23.61M 31m16s
9 GPT-5.6 Sol 41.48% $3.40 $4 / $20 N/A 43m32s
10 Wonder 39.12% $8.74 $4 / $20 5.90B / 45.38M 43m08s
11 Claude Haiku 5.5 38.50% $1.86 $0.1 / $0.5 11.08B / 103.36M 39m58s
12 Grok 4.6 38.09% $2.45 $2 / $6 1.28B / 38.71M 40m10s
13 Gemini 3.8 Flash 35.76% $2.98 $1.5 / $7.5 3.84B / 38.76M 22m00s
14 Grok 4.7 34.83% $7.12 $2 / $6 2.84B / 40.44M 40m34s
15 Ember-1 31.76% $10.99 $3 / $15 11.35B / 48.81M 1h18m
16 MiMo V2.6 Flash 31.72% $0.12 $0.14 / $0.28 4.92B / 66.88M 1h41m
17 DeepSeek V4.1 Flash 29.65% $0.29 $0.3 / $1.2 4.41B / 70.92M 1h07m
18 Step 5 Preview 29.58% $1.78 $1 / $2.7 8.89B / 56.53M 1h07m
19 GPT-6 Luna 28.38% $0.11 $0.1 / $0.5 1.77B / 29.98M 30m15s
20 Gemini 3.1 Pro Preview (02/26) 27.38% $2.56 $2 / $12 1.99B / 16.33M 18m03s
21 GPT-5.6 Luna 25.30% $0.26 $0.2 / $1.2 N/A 50m00s

Motivation

AI systems have tremendous potential to advance science. Unlike data from enterprise contexts, scientific data can be collected without coordinating large groups of people. Because ground truth verification in science domains is less dependent on human authorship, science evaluations also allow us to characterize superhuman capabilities. In BioMysteryBench, for example, a subset of tasks is explicitly not solvable by human experts. Many science tasks require the same underlying skills as coding and math tasks, but additionally require niche domain knowledge and discernment. The ultimate promise of expanding the science capabilities of AI systems is in the material improvement of the human condition via fundamental breakthroughs and discovery. Here, we make an effort to monitor advances in these capabilities.

Most science benchmarks test recall. The three benchmarks in this index give a model a dataset, a simulated lab, or a terminal and score what it finds:

  • BioMysteryBench: investigating real biological datasets to answer open questions
  • MysteryMechanism: recovering hidden mechanisms by running experiments
  • Terminal-Bench-Science: research workflows across the life, physical, earth, mathematical, and engineering sciences

The Vals Science Index averages the three into one number, so models can be compared on research work as a whole rather than on one field or task format.

Leaders by Discipline

No one model leads every science

The index's top five, ranked among all 21 index models.

Component GPT-6 Astra GPT-6.1 Sol Claude Sonnet 5.5 Claude Opus 5.5 Gemini 4 Argon
Index #1 #2 #3 #4 #5
Benchmarks
BioMystery #3 #2 #1 #4 #6
Mechanism #1 #4 #3 #2 #5
TB Science #1 #2 #4 #3 #5
TB Science disciplines
Life #1 #3 #4 #2 #6
Physical #3 #4 #1 #2 #5
Earth #4 #1 #11 #6 #2
Math #1 #2 #3 #4 #5
Engineering #3 #2 #1 #5 #9
  • #1
  • Top 5
  • Top half
  • Bottom half

Vals Science Index vs Vals Index

The science order is not the Vals Index order

Leaderboard rank on the Vals Index (left) and the Vals Science Index (right), for the models in the top 10 of both.

Vals Index Vals Science Index #7 GPT-6 Astra GPT-6 Astra #1 #9 GPT-6.1 Sol GPT-6.1 Sol #2 #2 Claude Sonnet 5.5 Claude Sonnet 5.5 #3 #3 Claude Opus 5.5 Claude Opus 5.5 #4 #1 Gemini 4 Argon Gemini 4 Argon #5 #5 Wonder Full Wonder Full #6 #6 Claude Opus 5 Claude Opus 5 #7 #10 Wonder Wonder #10

Methodology

Formula

The Vals Science Index is an even weighting of three benchmarks: BioMysteryBench, MysteryMechanism, and Terminal-Bench-Science.

Science_Index = AVG(BioMysteryBench, MysteryMechanism, TerminalBenchScience)

Updates

10/16/2026

Released Vals Science Index v1: the average of BioMysteryBench, MysteryMechanism, and Terminal-Bench-Science.