An AI model just worked through a stack of 50 unpublished, research-level theoretical physics problems. It solved 49 of them.
That result comes from a new preprint called TheoPhysEval, and the model in question is GPT-6-Astra. With web search turned on, it scored 98%, which is the kind of grade that usually gets a graduate student a handshake and a heavier workload.
What the benchmark actually tested
The paper is titled “TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving.” It landed on arXiv on October 9, 2026.
The benchmark contains 50 theoretical physics problems that had never been published. They span 11 subfields, including fluid dynamics and quantum field theory.
The goal was to see what AI agents can do with no human help at all. Nobody nudged the models along or fixed their mistakes midstream.
Physicists from institutions such as ICTS and Google DeepMind led the work. The author list runs to 35 people from a range of research institutions.
The researchers tested off-the-shelf models, not custom-built research systems. They designed the criteria to approximate the skills an expert theoretical physicist would need.
The scoreboard
The study split testing into two tracks. The “realistic” track let models search the web, the way a working researcher would. The “isolated” track cut them off from the internet entirely.
On the realistic track, GPT-6-Astra solved 49 of the 50 problems. The cost of that run came to $116.28.
On the isolated track, GPT-6-Astra’s score dropped to 46 out of 50. Web access, in other words, was worth three problems.
Opus 5.5 posted 46 out of 50 on both tracks. Giving it the internet didn’t move its score in either direction.
Then there’s the holdout. One problem went unsolved by every model the researchers tested. Fifty questions, and one of them proved stubborn enough to stump the whole field.
When the student corrects the answer key
Perhaps the most striking finding is not a score at all. According to the study, the models’ work led to corrections of some of the human-written reference solutions.
The use of unpublished problems matters here too. Problems that have never appeared in print are harder to have memorized from training data, so a correct answer is more likely to reflect actual reasoning than recall.
Why this matters beyond the lab
The cost figure deserves attention as well. A $116.28 run on the realistic track is small next to the time a team of human experts would spend working through 50 research-level problems, though the paper’s scope is performance, not a full cost comparison.
There are caveats worth keeping in view. This is a preprint, which means it has not yet gone through formal peer review. The benchmark is also 50 problems, a meaningful but finite sample of what theoretical physics involves.
And solving a well-posed problem is not the same as choosing which problem to solve. The benchmark measures autonomous problem solving, not the creative work of framing new questions, which remains a core part of what theoretical physicists do.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.