MIT Study: OpenAI Models Now Crack a Third of New Theory Problems

In sixteen months, OpenAI's models went from almost never solving a fresh theory problem to solving about a third of them. That is the central finding of "AI Disruption and Falling Costs in Theory Research", a paper dated October 13 by Gabriel Manso, Claudio Mayrink Verdun and Neil Thompson of MIT FutureTech. The researchers drew questions from 254 recent arXiv papers in seven domains spanning mathematics, physics and economics. On those questions, the share where a model reached the same conclusion as the human authors rose from 2.8% with o1 to 34.6% with GPT-5.5. The authors put the rate of improvement at 22 percentage points a year.

arXiv is the preprint server where physicists, mathematicians and a growing share of economists post work before, or instead of, journal publication. o1 was OpenAI's first so-called reasoning model, trained to work through a problem step by step before answering. The study covers model releases from December 2024 to April 2026. Coverage of the paper by iMasters and Superpower Daily lists six of the fields tested: probability, statistics, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.

The price is the part that should get the attention of anyone who funds research. The paper puts GPT-5.5's cost at about $5 per attempt, or $15 per success. By our arithmetic, $5 divided by a 34.6% hit rate comes to about $14.45. The per-success figure is the cost of the roughly three tries it takes to land one correct answer. The authors open their paper by noting that "a theoretical paper in mathematics or a related quantitative science represents a costly investment of expert time." A postdoc's afternoon costs a good deal more than $15.

Most of the study's design goes into keeping the models from cheating. Benchmarks built from published material tend to leak into the next generation of training data, which makes a high score hard to tell apart from good recall. To avoid that, the team took problems from papers posted between December 2025 and May 2026, after the models' training cutoffs, and blocked internet access during testing. Answers were then checked against solutions the models could not have seen.

The study has limits. Every model tested came from OpenAI, so the paper tracks one company's progress and does not compare labs. According to iMasters, the lineup was o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5. The problems were theory questions, so no experiments or simulations were involved. The scoring counts whether a model reached the same conclusion as the paper's authors. A matching final answer is strong evidence that a model is useful, but it does not show that every step of the reasoning behind it holds.

That gap has already caused trouble at the frontier. iMasters reports that in September OpenAI announced an unreleased model had solved the Navier-Stokes problem, which concerns the equations governing fluids in motion. The existence and smoothness of solutions to the Navier-Stokes equations is one of the Clay Mathematics Institute's Millennium Prize problems. According to iMasters, a mathematician at NYU who had been working on the same problem with an Anthropic-affiliated academic accused OpenAI of moving in on their work. The outlet also notes that OpenAI withdrew several solutions from its published results after they failed closer checking. One answer to the checking problem is machine-verified proof, which is the approach behind Claude agents formalizing Fermat's Last Theorem in Lean. OpenAI's own push on hard mathematics includes the Astra model we covered in August.

The MIT authors read their results broadly. They describe "a rapidly evolving shift in science in which some areas become substantially more cost-effective, making powerful theoretical tools available to a wider range of scientists." If the trend held at the same slope for another year, the share would pass half, by our arithmetic. Benchmark curves rarely stay straight, though. The more concrete change is already in the data: for about $15, a model can now return a correct answer to roughly one in three new research-grade theory problems in these fields.