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

A new benchmark called TheoPhysEval tested GPT-6-Astra on 50 unpublished theoretical physics calculations. The model solved 49 of them at a cost of $2.40 per problem, performing at a level comparable to leading specialists in the field.

Notably, in a few cases where the model appeared to fail, the experts' own reference solutions were found to contain errors. This suggests the benchmark itself may have had flaws, and the model's actual performance could be even stronger than the raw score indicates.

The result is significant because it demonstrates that frontier LLMs can now tackle open-ended, unpublished research problems in a specialized scientific domain, not just standardized or well-documented tasks. The per-problem cost of $2.40 is also striking, as it undercuts the cost of human expert time by orders of magnitude.

If these results hold up under broader scrutiny, they could accelerate research workflows in theoretical physics and related fields, where routine but demanding calculations currently consume substantial researcher time. The benchmark also highlights an emerging challenge: as AI systems approach or exceed expert performance, the quality of reference solutions and evaluation methodology becomes a critical bottleneck.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.