OpenAI is making an unusually explicit bet on scientific research. The company announced Kevin Weil, formerly Chief Product Officer, now leads a new division called OpenAI for Science. His mission, according to a post on his X account: building "the next great scientific instrument."
The timing coincides with GPT-5.2's release in December 2025, which OpenAI positions as its strongest model for math and science work. On the FrontierMath benchmark, which tests expert-level mathematics, GPT-5.2 Thinking solved 40.3% of problems, a new state of the art. On GPQA Diamond, covering graduate-level physics, chemistry, and biology, GPT-5.2 Pro hit 93.2% accuracy.
But the benchmark numbers matter less than a specific claim buried in OpenAI's announcement.
The proof that wasn't supposed to exist
OpenAI says GPT-5.2 Pro helped resolve an open research problem in statistical learning theory. The question, documented in a new paper, asks whether model performance reliably improves as more data is added. Researchers reportedly asked the model to solve it directly, without providing a proof outline. The model proposed a solution that was subsequently verified by human experts.
This follows earlier work from August 2025, when Microsoft researcher Sébastien Bubeck gave GPT-5 Pro an open problem in convex optimization. The model produced a new, tighter inequality that Bubeck verified as mathematically valid.
OpenAI is careful to frame these as collaborative achievements rather than autonomous breakthroughs. "These systems are not independent researchers," the company states. "Expert judgment, verification, and domain understanding remain essential."
What Weil actually wants to build
Weil's background makes the appointment interesting. Before tech, he studied high-energy physics at Stanford after graduating from Harvard with degrees in physics and math. At OpenAI, he said he'll work with the synthetic data team to "learn the craft" of AI research firsthand.
The new division has already hired its first researcher: Alex Lupsasca, a theoretical physicist from Vanderbilt who works on black hole physics. He'll maintain his academic position while guiding how frontier models approach complex scientific reasoning. In an Axios interview, Weil said GPT-5 can tackle astrophysics problems that would take talented graduate students days to approximate.
Weil is hiring what he calls a "small team" of academics who are "completely AI-pilled" and also "great science communicators." The marketing language aside, the approach suggests OpenAI wants researchers who can translate between model capabilities and actual scientific practice.
The usage numbers that matter more
OpenAI's broader usage data tells a different story than the science headlines suggest. A September 2025 study analyzing 1.5 million ChatGPT conversations found roughly 30% of usage is work-related, with the rest personal. Writing dominates work tasks. Coding and complex scientific work remain niche activities.
ChatGPT now has over 800 million weekly active users, with the fastest growth coming from low and middle-income countries, where adoption rates are running 4x higher than wealthy nations according to an NBER working paper. The gender gap that skewed heavily male in early 2024 has closed to roughly 50-50.
None of which tells us how many researchers are actually using ChatGPT for serious scientific work. OpenAI hasn't released that breakdown.
The skeptic's take
The science push arrives as OpenAI faces new competition. DeepMind, Anthropic, and others are pursuing similar research acceleration claims. Having a dedicated science division with a former CPO at the helm is partly positioning, partly genuine capability building.
GPT-5.2 solving a theorem is notable. Whether it represents "mathematical discovery" or sophisticated pattern matching remains an open question that philosophers and AI researchers will argue about for years. What's harder to dispute: the model produced a valid proof that humans subsequently verified. That's useful regardless of how you categorize it.
OpenAI says its approach "combines two complementary beliefs": specialized scientific tools remain essential for precision, while scaling foundation models unlocks new reasoning abilities. The company is betting it can do both.
GPT-5.2 Thinking and Pro began rolling out to paid users in December 2025.
