Study finds best OpenAI models now solve about 35% of theory problems, up from 2%

In his AI Lab newsletter for Wired, Will Knight describes a talk at MIT by Neil Thompson, who studies progress in computing and artificial intelligence. Thompson's latest study, conducted with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun, suggests that after mathematics, AI may move into probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.

The headline result: between mid-2024 and mid-2026, the best AI models went from correctly solving about 2 percent of theoretical problems in these fields to about 35 percent. Knight adds that there are no signs the progress will slow down. Thompson says: "Within just a few years, we expect a lot of theory questions to be totally transformed in terms of AI's ability to do them."

The design is built to avoid contamination. The researchers took problems stated in research papers posted online between December 2025 and May 2026, then tested six OpenAI models on them: o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5. The solutions appeared in papers published after the training cutoff for GPT-5.5 and the preceding models, so the models had to work out answers themselves rather than rely on training data. Those papers served as ground truth. The researchers also blocked internet access, so the models could not look solutions up.

The scope is limited. The trio focused on fields where a model can reason its way to an answer without running simulations or experiments. Even so, the work suggests that the shock waves felt in math are likely to spread to other math- and theory-heavy areas of science.

Knight sets this against the recent turmoil in mathematics. In September, OpenAI announced that an unreleased model had solved Navier-Stokes, a long-standing grand challenge concerning the equations of fluid motion. The episode was controversial: a mathematician at NYU, who had been working with another academic associated with Anthropic, accused OpenAI of gazumping them after learning of their progress. OpenAI and others have since announced hundreds more solutions to longstanding math problems, which Knight says triggered something of a crisis in the field. Some mathematicians have called for limits on AI use, warning that the process of solving problems is often as useful as the final answer.

Thompson says that, as with math, advances in other areas should be expected to have both positive and negative effects. He notes that theoretical advances sometimes drive practical advances in science, and suggests heavier use of AI may mean more progress happens that way. Some prominent AI researchers have founded new labs focused on AI for science, and the big AI labs are ramping up scientific research, exploring how AI could run real experiments alongside theoretical analysis.

The researchers say that theorists may no longer be required to solve key problems, but experts will still be needed to make sense of solutions and check that they are sound. Manso: "We are still going to need people to generate good ideas to connect the dots." He adds that for very complex problems AI might make mistakes, and that ways to verify results are needed. Knight notes in a parenthetical that OpenAI withdrew several solutions in its latest dataset after publication.

Key facts

Between mid-2024 and mid-2026, the best AI models went from correctly solving about 2 percent of theoretical problems in the studied fields to about 35 percent.
The fields named are probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics, all chosen because AI can work them out without simulations or experiments.
Six OpenAI models were tested (o1, o3, GPT-5, GPT-5.1, GPT-5.4, GPT-5.5) on problems from papers posted between December 2025 and May 2026, with solutions published after the models' training cutoff and internet access blocked.
Thompson expects a lot of theory questions to be 'totally transformed' within a few years; Manso stresses that people and verification methods will still be needed.

Why it matters

Mathematics has already been shaken by AI-generated solutions to long-standing problems, including OpenAI's Navier-Stokes announcement and hundreds of further claimed solutions. This study suggests the same pattern may reach other theory-heavy sciences. A rise from about 2 percent to about 35 percent in two years is the kind of curve Thompson points to when he expects theory questions to be transformed within a few years.

Who it affects

Theorists in probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics are the fields named. Mathematicians are the earlier case, some of whom have called for limits on AI use. Researchers in experimental work are less directly in scope, since the study covered only fields where AI can reason without simulations or experiments.

How to use it

This is a study report, not a product, so there is nothing to adopt directly. The practical takeaway is the evaluation method: test models on problems from papers published after their training cutoff, with internet access blocked, so that solving requires working out the answer. Manso's own advice points to pairing model output with ways to verify full solutions, not just final answers.

How solid is it

The method is careful about contamination: held-out papers from after the GPT-5.5 training cutoff, plus no internet access. The article does not say whether the study is peer-reviewed, preprinted, or where it is published. It does not give the number of problems tested or the number per field, and the 2 and 35 percent figures are aggregates across the fields. It does not say which specific model reached about 35 percent, only 'the best AI models'. Only OpenAI models were tested. The claim that progress will not slow is the article author's reading.

Risks and caveats

The study covered only fields where AI can figure things out without experiments, so it says little about experimental science. Manso warns that for very complex problems AI might make mistakes, and that a model can confidently assert a right answer that falls apart when the whole solution is examined. Knight notes OpenAI withdrew several solutions in its latest dataset after publication. Thompson expects both positive and negative effects, and some mathematicians warn that the process of solving problems is often as useful as the answer.

“Within just a few years, we expect a lot of theory questions to be totally transformed in terms of AI's ability to do them”

— Neil Thompson