GPT-6-Astra solves 49 of 50 unpublished physics problems in TheoPhysEval preprint
In brief
- GPT-6-Astra solved 49 of 50 unpublished theoretical physics problems with web search, per the TheoPhysEval preprint.
- GPT-6-Astra scored 46 of 50 without internet access; Opus 5.5 scored 46 on both tracks.
- ICTS and Google DeepMind physicists are among the authors; the paper hasn't been peer reviewed.
- One problem went unsolved by every model the researchers tested.
What the benchmark tested
The paper is titled "TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving." It contains 50 theoretical physics problems that had never been published, spread across 11 subfields (fluid dynamics and quantum field theory among them).
The point was autonomy. Per Crypto Briefing's report, nobody guided the models along or corrected their mistakes, and the researchers used off-the-shelf models rather than custom-built research systems.
Two tracks were used. A "realistic" track let models search the web, the way a working researcher would, while an "isolated" track cut off internet access entirely.
How the models scored
On the realistic track, GPT-6-Astra solved 49 of the 50 problems (a 98% score), and the preprint puts the cost of that run at $116.28. Without web access, its score fell to 46 out of 50.
Opus 5.5 posted 46 out of 50 on both tracks, according to the paper.
One problem went unsolved by every model the researchers tested.
The results also cut the other way. According to the study, the models' work led to corrections of some of the human-written reference solutions. The report doesn't give more detail than that.
What the paper doesn't claim
The authors set clear limits on what they measured. The paper's scope is performance, not a full cost comparison with human experts, so that $116.28 figure isn't a head-to-head against what a physicist would cost to do the same work.
TheoPhysEval measures autonomous problem solving. It doesn't measure the creative work of framing new research questions, which is a different skill and isn't something this benchmark tries to score.
And it's still a preprint. Until it clears formal peer review, every score above is the authors' own reported result, relayed here as Crypto Briefing reported it.
Frequently asked questions
What is TheoPhysEval?
TheoPhysEval is a benchmark described in an arXiv preprint posted October 9, 2026. It contains 50 previously unpublished theoretical physics problems across 11 subfields, including fluid dynamics and quantum field theory. It tests whether AI agents can solve them with no human help.
How did GPT-6-Astra and Opus 5.5 score on TheoPhysEval?
According to the preprint, GPT-6-Astra solved 49 of 50 problems with web search, at a run cost of $116.28, and 46 of 50 without internet access. Opus 5.5 scored 46 of 50 on both tracks. One problem went unsolved by every model tested.
Has the TheoPhysEval paper been peer reviewed?
No. The paper is a preprint and hasn't yet gone through formal peer review. Its scope is performance, not a full cost comparison with human experts, and it measures autonomous problem solving rather than framing new research questions.
