GPT-6-Astra Solves 49 of 50 Physics Problems at $2.40 Each

What happened

GPT-6-Astra solved 49 of 50 unpublished theoretical physics calculations on the TheoPhysEval benchmark, reaching accuracy comparable to leading specialists, according to a report from The Decoder. The reported cost works out at $2.40 per problem.

In several cases where the AI agent appeared to fail, the experts' own reference solutions were found to contain errors. That detail complicates a simple reading of the scorecard, because the answer key was not always the reliable baseline.

The researchers cited in the report argue that physics has not yet seen a breakthrough comparable to those in mathematics, because current agents lack what they describe as "taste". They still believe autonomous, top-level AI research could arrive soon, and they already expect effects on how young physicists are trained and on equal opportunity in the field.

Why it matters

The significance here lies in what the technology now makes possible. An agent that can work through unpublished, expert-level analytical problems at a reported cost of a few dollars each moves AI from assisting with well-defined tasks towards carrying out specialist analysis end to end. For organisations, that changes the economics of expert work, and it changes how quickly questions can be explored. Whether these results hold under independent scrutiny will matter a great deal, and the report itself suggests the evaluation process needs as much attention as the model.

The human side of the story is equally important. If AI takes over a growing share of routine expert calculation, the apprenticeship pathway that trains young physicists could be reshaped, and the question of who gets access to advanced research capability becomes a design choice rather than an accident. Any organisation that relies on expert judgement, from research labs to professional services, will need to decide where that judgement is captured, how it is checked, and how new practitioners will still learn the craft.

By the numbers

• 49 of 50 unpublished theoretical physics calculations solved by GPT-6-Astra on the TheoPhysEval benchmark
• $2.40 reported cost per problem, according to the headline reporting

The Renascence take

Most readers will fixate on the headline score. The more useful signal is the finding that expert reference solutions contained errors, which shows the benchmark itself is a service artefact that needs quality assurance.

The uncomfortable lesson is that a 49-out-of-50 score is only as trustworthy as the answer key behind it, and many organisations would never audit theirs. The real design challenge is "taste": the judgement that tells a good answer from a plausible one, and that is rarely written down. Customer-obsessed operators should map where that judgement lives in their own services, build structured review loops around AI output, and treat accuracy claims as a starting hypothesis rather than a verdict.