WIRED reports that the best-performing models in a six-field theoretical science benchmark solved about 35% of its problems around mid-2026, up from about 2% around mid-2024. The test covered theory problems that did not require simulations or experiments, so the result describes performance on that benchmark—not a general measure of AI’s scientific ability.
What the reported increase means
The two rounded rates differ by about 33 percentage points (35% minus 2%). That calculation conveys the size of the reported change; it does not show that performance improved evenly across the fields or across scientific research as a whole. WIRED’s account of the study attributes the research to Neil Thompson, Gabriel Manso, and Claudio Mayrink Verdun.
What the benchmark tested
WIRED says the researchers tested OpenAI’s o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5 on problems drawn from research papers posted online between December 2025 and May 2026. The six fields were probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. Internet access was blocked during the tests, and the problems were selected so they could be addressed without simulations or experiments.
Why the result has a defined boundary
A theoretical problem can be worked through without running an experiment; many scientific questions cannot. Because this benchmark excluded simulation- and experiment-dependent problems, its rates do not measure how the models perform in experimental science.
Checking the solution matters, too. WIRED reports that OpenAI withdrew several solutions from a dataset after publication. Gabriel Manso, a member of the study team, emphasized both the need for people to develop useful ideas and the need to verify a model’s full solution. NeoTeo previously covered scrutiny of a separate OpenAI mathematics claim in its report on the Navier–Stokes challenge.
BootLoops illustrates a different research workflow
A separate effort described by Science used Claude with BootLoops, a tool suite for finding mathematically structured problems across scientific literature. The report says the software was used to solve 30 examples of a class of integrals, including 15 described as previously unsolved. The wider work produced 36 preprints after three months of work.
Those projects are not another measurement of the six-field benchmark: they describe AI-assisted research across different problems and fields. Science also stresses the role of researchers in finding worthwhile questions and judging the results. The benchmark’s rising score and BootLoops’ research outputs therefore describe different kinds of work, with human expertise involved in both.
