GPT-6-Astra solves 49 of 50 unpublished theoretical physics problems OpenAI

Physicists from 19 institutions built TheoPhysEval, a set of 50 unpublished research-level theoretical physics problems, and tested AI agents on it. GPT-6-Astra running in Codex solved 49 of 50, and in several cases its answers were right where the experts' own reference solutions were wrong. The paper went up on arXiv on October 9, 2026.

The authors first marked Astra wrong on several problems. On closer examination with the experts, they found subtle errors in their own original solutions, and in all of those cases Astra's answer turned out to be correct.

What it is

• The problems. Experts wrote the hardest problems they could in their own fields. A world expert would solve a typical one in about a day or less; experts from neighbouring fields would take much longer. None of the problems are public.
• The checking. Every answer is reduced to a number and checked automatically against the reference.
• Two tracks. On the realistic track agents may search the web and write scripts. On the isolated track they have no internet.
• The one miss. The only problem Astra failed on the realistic track is an inverse problem: find the differential equation behind a given curve. The authors say its solution "involves some play on the terms in the prompt". No model solved it.

Results

Accuracy barely changed without web access. The authors read this as a sign that the models have memorized large parts of the physics literature.

Price

Evaluation cost for all 50 problems, in USD, from the paper's token usage:

That puts Astra under USD 3 per problem on average.

Sources

• TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving (arXiv:2610.16733)