GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece

GPT-6-Astra solved 49 of 50 unpublished theoretical physics problems on a benchmark called TheoPhysEval, according to reporting by The Decoder. The agent matched leading specialists in accuracy, at a cost of about $2.40 per problem. In several cases where the agent appeared to fail, the specialists' own reference solutions turned out to contain errors.
Despite that result, the researchers say physics has not seen a breakthrough of the kind that has occurred in mathematics, attributing the gap to the agents lacking what they call "taste." They still expect autonomous, top-level AI research could arrive soon, and they already point to consequences for how young physicists are trained and for equal opportunity in the field.
Why it matters: A near-perfect score on unpublished research problems, at a few dollars each, is the kind of result that prompts labs to reconsider how much routine theoretical calculation they hand to agents. The reported errors in expert reference solutions also suggest evaluation itself needs scrutiny before anyone treats a benchmark score as settled. Researchers' caution about missing "taste" is a reminder that solving well-posed problems is not the same as choosing which problems are worth solving.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
On the TheoPhysEval benchmark, GPT-6-Astra solves 49 of 50 unpublished theoretical physics calculations as accurately as leading specialists. In some cases where the agent seemed to fail, the experts' own reference solutions turned out to contain errors. But the researchers say physics hasn't had a breakthrough like the ones in mathematics, because the agents lack "taste." They still think autonomous, top-level AI research could arrive soon. They already see consequences for how young physicists are trained and for