Frontier AI Models Now Solve 35 Percent of Hard Theory Problems

The best AI models went from solving about 2 percent of theoretical problems in six scientific fields in mid-2024 to about 35 percent of those problems by mid-2026, according to a new study by Neil Thompson's lab at MIT. The finding, presented by Thompson in a talk at MIT last week, covers probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.

The stakes are direct. Mathematicians are still absorbing OpenAI's announcement last month that an unreleased model solved Navier-Stokes, a grand-challenge problem concerning the equations of fluid motion. Thompson's work suggests the same shock waves will now reach every math-heavy corner of scientific inquiry, not just pure mathematics.

"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," Thompson said.

Who ran the study, and how?

Thompson studies progress in computing and artificial intelligence; his lab examines how advances in computer science affect society and innovation. He conducted the new study with Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist.

The method was strict about contamination. The trio took problems stated in research papers posted online between December 2025 and May 2026, then tested whether several OpenAI models—o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5—could solve them. The papers containing the solutions appeared after the training cutoff for GPT-5.5 and preceding models, so no model had seen the answers in its training data. The researchers also blocked internet access, which prevented the models from looking up solutions online.

The published papers served as ground truth: each solution the models produced could be checked against work done by humans.

What does the 2-to-35 percent jump actually measure?

It measures correct solutions to theoretical research problems in the six fields named above, across the two-year window. Thompson's case, made over the course of half an hour at MIT, is that several scientific fields are about to be upended—and that the trajectory, not the absolute number, is the signal. The study reports no sign that the progress will slow.

The scope has limits. The researchers focused on fields in which AI models could reason to answers without running simulations or physical experiments. That boundary matters for interpreting the result: the study covers theory, not laboratory science. Even so, the work indicates that the disruption now visible in mathematics is likely to spread to other theory-heavy areas of research.

Why does this matter for mathematics?

The math community provides the template for what comes next. Last month, OpenAI announced that an unreleased model had solved Navier-Stokes, a long-standing grand challenge concerning the equations that explain fluid motion. The episode turned controversial when 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. The announcements have triggered what the source describes as something of a crisis in the field. Some mathematicians have called for limits on the use of AI, arguing that the process of solving a problem is often as useful as having the final answer.

Thompson's study extends that debate from mathematics to physics theory and economic theory. If frontier models can solve a third of published theoretical problems today, and the curve has not bent, the economics of theoretical research change: problems that once defined careers become tasks.

What are the upsides the researchers see?

Thompson expects advances in other areas to carry both positive and negative effects, as in mathematics. He notes that theoretical advances sometimes drive practical advances in science, and suggests that greater use of AI may mean more progress happens that way.

The commercial and institutional signals point the same direction. Some prominent AI researchers have already founded new labs focused on AI for science, and the major AI labs are all ramping up scientific research. These efforts explore how AI could be empowered to conduct real experiments alongside theoretical analysis—an extension beyond the theory-only scope of Thompson's benchmark.

Who still needs to be in the loop?

The researchers are explicit that automation of solutions does not eliminate the need for experts. AI may mean theorists are no longer required to solve key problems, but the study's authors say experts remain necessary to make sense of solutions and confirm they are sound.

"We are still going to need people to generate good ideas to connect the dots," Manso said. He added that for very complex problems, it is possible that AI might make mistakes.

The record supports the caution. OpenAI withdrew several solutions in its latest dataset after publication. "We also need ways to verify what was done, because it's very easy for a model to tell you 'this is the right answer,' and when you look at the entire solution, you start to see maybe it is not," Manso said.

The open question is verification at scale. If models keep climbing from 35 percent toward the bulk of solvable theory, the bottleneck shifts from producing answers to checking them—and the institutions that fund and publish theoretical science will have to decide how fast to follow.

This story is based on reporting by Will Knight.