OpenAI's GPT-5.2 derives new theoretical physics result

Standard physics textbooks have long claimed that a specific particle interaction simply cannot happen. OpenAI’s GPT-5.2 just proved the textbooks wrong.

In a new preprint co-authored with researchers from the Institute for Advanced Study, Harvard, Cambridge, and Vanderbilt, OpenAI revealed that a certain gluon interaction, widely assumed to have a zero scattering amplitude, is actually nonzero under specific kinematic conditions.

Even wilder than the physics itself is how the result was discovered: an AI model conjectured the formula, and an internal reasoning model spent 12 hours proving it.

Smashing through textbook assumptions

The paper zeroes in on scattering amplitudes, the mathematical engine physicists use to calculate the probabilities of particle collisions. For gluons, the carriers of the strong nuclear force, amplitudes at tree level often collapse into surprisingly simple equations.

However, standard textbook arguments have treated one specific setup as nonexistent. When one gluon has negative helicity and the remaining n-1 gluons have positive helicity, generic assumptions dictate the amplitude vanishes.

The researchers showed that this assumption fails on a precise slice of momentum space known as the half-collinear regime. On this slice, the amplitude is strictly nonzero.

The real breakthrough was the methodology. The human authors calculated base cases by hand up to six particles. Beyond that, the complexity of the Feynman diagrams exploded super-exponentially into an unmanageable mess.

Enter GPT-5.2 Pro. The model drastically simplified the hideous hand-calculated expressions, spotted the underlying pattern, and posited a general formula for any integer n.

Then, an internal scaffolded version of GPT-5.2 spent roughly 12 hours reasoning through the problem. It produced the exact same equation alongside a formal proof. The team subsequently verified the math analytically using standard Berends-Giele recursion relations and the soft theorem. The authors note the approach has already been extended to gravitons.

Why it matters

We think this is a genuine milestone for AI in science. We have seen models summarize literature, write boilerplate code, and predict protein folds. But high-energy theoretical physics requires abstract mathematical intuition, something many assumed LLMs could never grasp.

For theoretical physicists and mathematicians, this marks the arrival of AI as a legitimate research collaborator. Spotting clean formulas inside messy algebra has always been an art. GPT-5.2 proves that pattern recognition in symbolic physics can be delegated to machines, freeing researchers to focus on interpreting physical reality.

OpenAI’s execution here was brilliant because it established an ironclad verification loop. Instead of letting the model hallucinate speculative theories, they fed it rigorous base cases, had it generate a conjecture, ran a scaffolded 12-hour reasoning proof, and cross-checked everything against proven physics theorems.

The caveat is accessibility. You cannot just prompt ChatGPT to revolutionize quantum field theory today; the heavy lifting was done by a custom, internally scaffolded system run across half a day. Still, this paper changes the competitive landscape for frontier models, proving that long-horizon reasoning can produce novel, peer-reviewable science rather than just better chatbot replies.