An AI agent called GPT-6 Astra reached the correct final answer on 49 of 50 research-level theoretical-physics problems that had never been published, according to TheoPhysEval, a preprint posted on October 9, 2026, by Ramesh Chandra and colleagues. Because the problems were new, their answers could not be looked up online. The result, and what it might mean for people who choose to become physicists, came up on an episode of Moonshots with Peter Diamandis recorded on October 16 and published the next day.
Diamandis, who founded XPRIZE and hosts the show, presented it as a sequel. The week before, he said, OpenAI had "dropped 722 math proofs" on GitHub, and the panel had long expected physics to come next. OpenAI's math repository lists 719 manuscripts grouped into 372 result families, at different stages of verification. Not all of them have Lean proofs, versions of the argument written in a formal language that a computer can check step by step. The company announced the release on October 6.
What the physics test measured
The preprint tested five AI agents, meaning language models given tools and time to work through a multistep task on their own. Experts supplied 50 calculations across 11 physics themes, either problems taken from their own research notes or adaptations of published calculations. An automated grader checked each final numerical answer. Each agent got one run per track, usually with an eight-hour limit per task.
In the track with web access, Astra scored 49 of 50, followed by Opus with 46, Gemini with 34, DeepSeek with 27 and GLM with 21. Without web access, Astra and Opus tied at 46, with Gemini at 31, GLM at 26 and DeepSeek at 25. The authors note that single runs and uneven coverage of physics fields limit how precise these scores are.
Diamandis described the trade as a world-class physicist's day per problem against roughly $3 of computing. The preprint's figures put it a little lower: an estimated $116.28 and $95.20 for Astra's two complete 50-problem runs, or about $1.90 to $2.33 a problem.
He also passed on what he called an "interesting detail": Astra was first reported to have missed one of the 50, but the published answer to that problem turned out to be wrong, and Astra had it right. The preprint gives a more mixed account. It says several reference solutions were corrected after researchers examined what looked like model errors, and it still reports Astra at 49 of 50, with one inverse problem left unsolved by every agent.
Verifiable domains
When Diamandis asked what the result showed, the discussion held that in verifiable domains, where a proposed answer can be checked against a right one, as in mathematics and much of theoretical physics, AI is now simply better than people. It cited a no-lookup score of 42 out of 50. The preprint puts Astra's no-web score at 46. The argument was that a tool able to verify anything verifiable lets people build much more complex systems and processes.
Salim Ismail, founder of Open ExO, applied that reasoning to business tasks with clear outcomes, such as the know-your-customer checks banks run on new clients. In a pilot he is running, he said, 22-step workflows have dropped to four steps, and he expects that kind of change to make organizations "10x more efficient."
Professor level, and a half-year forecast
Diamandis then asked how hard the 50 problems really were, compared with the hardest open problems of the kind associated with the Millennium prizes. The answer in the discussion placed them at roughly professor level, on the grounds that a professor could check each one within a day. It went on to say physics' biggest open questions, such as quantum gravity, the role of time in quantum mechanics and a unified field theory, would probably not fall to this kind of problem-solving but to another route, and within no more than half a year. Diamandis's response was "oh my God."
Ismail recalled that Eric Schmidt, speaking on stage at Diamandis's A360 event a couple of years ago, had imagined a time when every lab in the world would have the equivalent of the world's best physicist. Asked when, Schmidt said five to six years. "And now it's like two years. And here we are," Ismail said, adding that every lab now has access to "incredible competence at a few bucks an hour."
Who asks the next question
The conversation turned to XPRIZE, the competition foundation, and whether AI could check and extend the science its prizes have produced. Anousheh Ansari, the foundation's CEO, agreed that it "definitely can help the validation process," then raised a worry. If AI is "bulk solving" the problems people have already defined, she said, young people choosing a field to study "probably will not be very excited to go in a field where AI can do the job better." The article about the test, she noted, said AI does not yet have the curiosity to identify new problems, "and I don't know if it will ever get there." Without a new generation of mathematicians, physicists and astrophysicists finding new challenges and mysteries, she asked, "how are we going to advance?"
The preprint's authors draw a similar line. They separate getting calculations right from choosing productive research directions, and they write that in their experience human experts still supplied that direction.
The reply in the discussion was that humans can still steer, because the question is what is good for humanity. The proof of Fermat's Last Theorem was fantastic mathematics, the argument went, but made little difference to ordinary people. Because there is only one physics, something must connect its laws, and new models will help find it. After that, the value lies in applying the framework: to build flying cars, to explore the universe, to study biology. On that view, the work shifts from mapping the contours of the universe to using what is known to make a difference in it.