What was solved is a calculation, not an open grand problem
The Decoder reported on October 15 that 35 researchers released TheoPhysEval. The institutions include the International Centre for Theoretical Sciences, Harvard, KU Leuven, and Google DeepMind. The publish time is 15:24 UTC, 11:24 p.m. in Beijing. The benchmark has 50 unpublished theoretical-physics problems with answers that can be checked. The authors say a leading expert in the subfield, without AI, would need about a day or less for a typical problem.
GPT-6-Astra, in OpenAI's Codex, got 49 of 50 right, or 98 percent. The body says the average cost was under $3 per problem and just under $116 for the whole set. The page title says $2.40 apiece. Both figures stay. They are not collapsed into one number. This is a news report, not the paper. The report says the dataset is on Hugging Face and does not give the address.
[1]
Without the web, the score is still 92 percent, and one problem stopped every model
The authors call the items turn-the-crank problems: known methods applied to a slightly changed question. Major open problems are left out on purpose. Contributors often changed a few values in a calculation from one of their own recent papers so the new answer could not be guessed from the old one. The authors say problems like these used to be a first step for PhD students.
With web search, Anthropic's Opus 5.5, in Claude Code, was second at 92 percent. Google's Gemini 3.8 Flash scored 68 percent. The open-weight models DeepSeek-V4.1-Flash and GLM 5.3 reached 54 and 42 percent. Without web search, Astra and Opus each scored 92 percent. Each agent ran twice, once with web search and once without it.
One problem stopped every model. It asked for an equation that produces a given curve. The authors admit it might seem unfair, and they still treat it as close to how researchers infer physics from imprecise data. On several other items, follow-ups found small errors in the experts' reference answers. After those were fixed, Astra's answer was correct every time.
[1]Taste is not quantified, and it is not a forecast
The authors write that some of the strongest current models may already have passed the best human researchers on concrete calculations with a checkable result, even in specialized fields. They call that a major milestone. That is their judgment. The line about the fluid-equation Millennium problem is attributed to OpenAI in the report. This article does not verify a solution.
The authors stress that the remarks on taste and breakthroughs cannot be quantified and are not a forecast of where AI is headed. They also expect the cost of harder problems to rise sharply. That is an expectation. A full run of this benchmark is still about $116.
[1]