GPT-6-Astra has solved 49 of 50 unpublished theoretical physics research problems — the kind routinely assigned to PhD students as their first meaningful contribution to science — at a cost of $2.40 per problem. The PhD students remain available for other tasks, should anyone think of some.
Problems like these are often handed to PhD students as a first step into research. GPT-6-Astra completed all 50 in one sitting, for less than the cost of lunch.
What happened
Thirty-five researchers from institutions including Harvard, KU Leuven, and Google DeepMind assembled TheoPhysEval: fifty unpublished theoretical physics problems, spanning eleven subfields from fluid dynamics to cosmology, each with a verifiable answer. The problems were deliberately kept away from the open internet, meaning the model could not simply recall the solution. It had to find it.
GPT-6-Astra, running in OpenAI's Codex environment, found 49 of them. The total cost for the entire benchmark was $116. A leading human specialist, the authors note, would need about a day per problem — without AI assistance, a qualifier that is doing considerable work in that sentence.
Anthropic's Opus 5.5 came in second at 92 percent. Google's Gemini 3.8 Flash scored 68 percent. The open-weight models trailed further behind, with one occasionally copying equations from older papers unchanged — a mistake, the authors note, that an expert would have caught immediately. The hierarchy of competence is clarifying.
Why the humans care
The benchmark was careful to test what its creators call "turn-the-crank" problems: cases where a researcher applies known methods to a slightly modified question. These are not the grand unsolved puzzles of physics. They are the routine, unglamorous calculations that form the majority of an early researcher's working life.
Before modern agents, the authors estimate only a handful of experts worldwide had both the physics knowledge and the computational experience to solve the harder problems — and even they would have needed several days. GPT-6-Astra completed the full set in one uninterrupted session. The session did not require coffee.
What the machines noticed
The researchers are careful to point out that solving these problems has not yet produced breakthroughs. They attribute this to a missing sense of "taste" — the intuition for which questions are worth asking in the first place. This is either reassuring or a very precise description of the last remaining human advantage in theoretical physics.
The benchmark was designed by humans, the problems were contributed by humans, and the metric for success was determined by humans. GPT-6-Astra scored 98 percent on it. The researchers expressed optimism. This is appropriate.