An artificial intelligence model has worked through a set of 50 unpublished, research-level theoretical physics problems — and solved 49 of them.

The result comes from a new preprint, TheoPhysEval, and the model behind it is GPT-6-Astra. With web search enabled, the system scored 98% — the sort of grade that usually earns a graduate student a handshake and a heavier workload.

What the benchmark tested

The paper, titled “TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving,” appeared on arXiv on October 9, 2026. It introduces a benchmark of 50 theoretical physics problems that had never previously been published, drawn from 11 subfields ranging from fluid dynamics to quantum field theory.

The aim was to see what AI agents can achieve with zero human intervention. No one nudged the models along, and no one fixed their mistakes midstream. The “agents” label matters: these are systems that carry out multi-step tasks on their own — here, searching for information and working through to full solutions — rather than answering a single prompt in one shot.

Physicists from institutions including ICTS and Google DeepMind led the work, and the author list runs to 35 people across a range of research institutions. The team evaluated off-the-shelf models rather than custom-built research systems — meaning the results reflect what general-purpose models can do, not the output of tools specially engineered for physics — and designed grading criteria to approximate the skills an expert theoretical physicist would need.

The scoreboard

Testing ran along two tracks designed to isolate the value of outside information. The “realistic” track let models search the web, as a working researcher would, while the “isolated” track cut them off from the internet entirely.

On the realistic track, GPT-6-Astra solved 49 of the 50 problems at a total cost of $116.28. On the isolated track, its score slipped to 46 out of 50. Web access, in other words, was worth three problems.

Opus 5.5 scored 46 out of 50 on both tracks. Granting it web access made no difference in either direction.

Then there is the holdout: one problem that every model the researchers tested failed to crack. Across all fifty questions, a single one proved stubborn enough to stump the entire 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 in some of the human-written reference solutions.

The use of unpublished problems matters here as well. Questions that have never appeared in print are harder to have memorized from training data, so a correct answer is more likely to reflect genuine reasoning than recall.

Why it matters beyond the lab

The cost figure also deserves attention. 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 rather than a full cost comparison.

Caveats remain in view. This is a preprint, meaning it has not yet undergone 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. Both points leave open questions worth tracking: whether the results hold up under formal peer review, and whether any future run clears the single problem that stumped every model tested.