Near-expert performance in physics
OpenAI's GPT-6 Astra model solved 49 of 50 unpublished theoretical physics calculations on the TheoPhysEval benchmark, matching the accuracy of leading specialists. Each solution cost $2.40.
Errors found in reference solutions
In an interesting twist, some cases where the model appeared to fail turned out to involve errors in the experts' own reference solutions. This suggests the evaluation process tested human solutions as much as the model itself.
The missing 'taste'
According to the researchers, physics has not yet seen a breakthrough like those in mathematics. They attribute this to agents lacking 'taste'—the ability to choose which problems matter and set a creative direction.
Implications for training and opportunity
The team believes autonomous, top-level AI research could arrive soon. Consequences are already visible in how young physicists are trained and in equal opportunity within the field.
