GPT-6-Astra Solves 49 of 50 Theoretical Physics Problems at $2.40 Each
On the TheoPhysEval benchmark, the AI agent matched leading specialists on almost all unpublished calculations. Researchers say the missing ingredient is scientific taste, but warn that autonomous top-level research may arrive soon.
An artificial intelligence agent has solved 49 of 50 unpublished theoretical physics calculations at a cost of roughly $2.40 per problem, matching the accuracy of leading specialists in the field. The results, obtained on a benchmark called TheoPhysEval, mark a significant step for machine-driven scientific reasoning, though researchers caution that the achievement does not yet amount to a genuine breakthrough in physics.
The benchmark consisted of unpublished problems, meaning the agent could not rely on memorized solutions from its training data. In several cases where GPT-6-Astra appeared to fail, the experts' own reference solutions were later found to contain errors, suggesting the agent's answers may have been more reliable than the evaluation itself. The cost of $2.40 per problem stands in stark contrast to the time and expense of human specialist work, raising questions about how such tools might reshape research economics.
Despite the high success rate, the researchers behind the evaluation say physics has not experienced the kind of dramatic advances seen in mathematics, where AI systems have already contributed to notable discoveries. The reason, they argue, is that the agents lack what scientists call «taste» — the intuitive judgment that tells a researcher which problems are worth pursuing, which approaches are promising, and when a result feels significant. Without that capacity, the agent can compute answers but cannot set its own research agenda.
That limitation may not hold for long. The same researchers believe autonomous, top-level AI research could arrive soon, and they already see consequences for how young physicists are trained. If an AI can solve most theoretical problems at low cost, the value of traditional problem-solving drills may decline, shifting the emphasis toward skills that machines cannot yet replicate: framing questions, interpreting results, and exercising scientific judgment.
The findings also raise concerns about equal opportunity in the field. If access to powerful AI research agents becomes uneven, institutions and countries with the resources to deploy them could pull ahead, while others fall behind. The low per-problem cost of $2.40 suggests the technology could eventually be widely available, but the infrastructure and expertise needed to use it effectively may remain concentrated.
TheoPhysEval is designed to test theoretical physics calculations that have not been published, making it a harder measure than benchmarks based on existing literature. The agent's performance on 49 of 50 problems indicates that it can handle complex derivations and reasoning tasks that were previously considered the domain of highly trained human experts. The discovery of errors in the reference solutions adds a layer of complexity to the evaluation, since it suggests that even expert benchmarks may need revision when AI systems perform at this level.
Researchers stress that solving problems is not the same as making discoveries. In mathematics, AI has helped identify new patterns and assist with proofs, but physics requires connecting abstract calculations to the physical world, designing experiments, and distinguishing between mathematically valid results and physically meaningful ones. The absence of «taste» means the agent cannot yet decide which calculations matter or how to interpret them in the context of a broader theory.
Still, the trajectory is clear. The combination of high accuracy, low cost, and rapid improvement suggests that AI agents will increasingly be used as tools in theoretical research, handling routine calculations while human physicists focus on the conceptual and strategic work. How quickly that division of labor evolves — and whether it leads to a genuine breakthrough in physics — remains an open question, but the researchers behind TheoPhysEval believe the day of autonomous, top-level AI research is approaching faster than many in the field expect.
6
