OpenAI published a preprint on arXiv on February 13, 2026, claiming its GPT-5.2 model conjectured a closed-form formula for a class of gluon interactions that textbooks have treated as impossible for decades. The paper, co-authored with researchers from Harvard, Cambridge, the Institute for Advanced Study, and Vanderbilt, argues that so-called "single-minus" gluon tree amplitudes don't actually vanish under certain momentum conditions.
The result hasn't been peer-reviewed yet. But the author list is hard to dismiss: Andrew Strominger from Harvard, David Skinner from Cambridge, Alfredo Guevara from IAS, and Alexandru Lupsasca, who holds positions at both Vanderbilt and OpenAI. Kevin Weil, OpenAI's Chief Product Officer, is also credited on behalf of the company.
What the textbooks got wrong (sort of)
Scattering amplitudes tell physicists the probability that particles will collide and produce a specific outcome. For gluons, the particles carrying the strong nuclear force, most of these calculations simplify into compact expressions at "tree level," the simplest tier of quantum field theory diagrams. One configuration was always skipped: when a single gluon has negative helicity and all the rest have positive helicity. Standard power-counting arguments said that amplitude had to be zero.
Turns out the standard argument comes with fine print. It assumes "generic" particle momenta, meaning no special alignment between the directions and energies of the particles. The OpenAI blog post accompanying the paper explains that in a precisely defined region of momentum space called the "half-collinear regime," where gluon momenta obey a specific alignment condition, the old reasoning breaks down. The amplitude doesn't vanish. It's not zero.
This isn't a rewrite of quantum chromodynamics. It's a loophole, and it only applies on a thin mathematical slice of momentum space. The textbooks weren't wrong so much as incomplete.
So what did the AI actually do?
This is the part that matters, and it's more nuanced than the headlines suggest. Human physicists on the team calculated the amplitudes by hand for up to six particles. The resulting expressions were, by all accounts, a mess. Equations 29 through 32 in the paper span dozens of terms, with complexity growing superexponentially as you add particles. Nobody could see a pattern in that.
GPT-5.2 Pro simplified those expressions into more compact forms. From the simplified cases, it spotted a regularity and proposed a general formula, Equation 39, valid for any number of gluons. "In short order, GPT-5.2 Pro suggested a beautiful and general formula for arbitrary n," Kevin Weil wrote on X, "but couldn't prove it."
That last part is important. The model conjectured but couldn't close the deal. An internal scaffolded version of GPT-5.2, given more compute and time, spent roughly 12 hours reasoning through the problem and produced a formal proof. The human authors then verified it against the Berends-Giele recursion relation and Weinberg's soft theorem, standard consistency checks in the field.
Pattern recognition feeding into conjecture, followed by extended machine reasoning, followed by human verification. That's the actual workflow. Not "AI discovers new physics" but something more like "AI proposes formula that humans confirm is correct."
The credibility question
Nima Arkani-Hamed, a professor at the Institute for Advanced Study who is not a co-author, offered a quote for OpenAI's blog post that reads like measured enthusiasm rather than hype. He called the scattering processes "something I've been curious about since I first ran into them about fifteen years ago," and noted that simple formulas in this area of physics often point toward deeper, undiscovered structures.
Nathaniel Craig, a physicist at UC Santa Barbara, went further, calling it "clearly journal-level research." That's a strong endorsement, though it's worth noting both quotes appeared in OpenAI's own announcement.
Skeptics on Hacker News and Reddit have pushed back, calling the AI's contribution "brute force pattern matching" and pointing out that humans still had to define the half-collinear regime for the model to work within. There's also the CritPt problem: GPT-5.2 initially scored 0% on this research-level physics reasoning benchmark, though a later evaluation with maximum reasoning effort brought it to 11.6%. A model that can conjecture a formula that stumped physicists for decades but struggles with structured physics exams is, at minimum, a strange object.
The paper itself is on arXiv and hasn't gone through journal peer review. OpenAI says it's been submitted and welcomes community feedback. The team claims to have already extended the approach from gluons to gravitons, with follow-up publications expected.
