Vals Science Index: GPT-6 Astra Leads 21 Models With 65.09%
OpenAI's GPT-6 Astra finished more than 5 points clear of GPT-6.1 Sol on the first Vals Science Index, which averages three benchmarks at equal weight.
Vals AI published the first Vals Science Index on October 16, which ranks AI models on scientific research ability, and OpenAI's GPT-6 Astra took first place. Twenty-one models took part in the initial evaluation, according to aitimes.com.
GPT-6 Astra scored 65.09% overall, more than 5 percentage points ahead of second place. GPT-6.1 Sol ranked second at 59.63%, followed by Anthropic's Claude Sonnet 5.5 at 59.12% and Claude Opus 5.5 at 58.65%. Google's Gemini 4 Argon was fifth at 55.36%.
The scoring splits differently by task. GPT-6 Astra ranked first on the hidden-mechanism experiment benchmark and on real science research tasks, but third on biology data analysis. Claude Sonnet 5.5 led biology data analysis and also led in physics and engineering research tasks, while GPT-6.1 Sol scored best in earth science research tasks.
The index averages three benchmarks with equal weighting: BioMysteryBench, developed by Anthropic; MysteryMechanism, developed by Vals AI; and Terminal-Bench Science.
Different leader on general work
On Vals AI's separate general-work index, Gemini 4 Argon ranked first with 68.90%, while GPT-6 Astra placed seventh with 63.13%.
In August, Vals AI raised $40 million in a round led by Andreessen Horowitz (a16z) at a $400 million valuation, the outlet reported. Vals AI was founded in the US in 2024.
Quick answers
Which model topped the first Vals Science Index?
OpenAI's GPT-6 Astra ranked first with 65.09%, more than 5 percentage points ahead of second place.
How is the Vals Science Index scored?
It averages three benchmarks at equal weighting: BioMysteryBench, MysteryMechanism and Terminal-Bench Science.
How many models were evaluated?
Twenty-one AI models took part in the first Vals Science Index evaluation.