THE CRUNCH
On the TheoPhysEval benchmark, GPT-6-Astra solved 49 of 50 unpublished theoretical physics calculations as accurately as leading specialists, at a reported cost of about $2.40 per problem. In some cases where the agent appeared to fail, the experts' own reference solutions turned out to contain errors.
The researchers behind the evaluation still see a gap between the model and genuine scientific discovery. Physics, they say, has not yet had a breakthrough comparable to those seen in mathematics, because the agents lack what they call "taste". Even so, they think autonomous, top-level AI research could arrive soon.
They already see consequences closer to home, particularly for how young physicists are trained and for equal opportunity across the field.