GPT-6 Astra helps crack a 50-year-old lattice physics problem
GPT-6 Astra turned physicist Nathaniel Craig's strategy into a working proof, addressing a 25-year-old phase problem blocking full Standard Model lattice simulations.

Why it matters
- GPT-6 Astra helped produce a proof addressing a 50-year-old problem in particle physics, presented in a preprint and on the OpenAI blog.
- The overlap fermion phase problem had gone unsolved for over 25 years, blocking full Standard Model lattice simulations.
- The result extends a 2026 run of AI math advances, including a 166-page OpenAI Navier-Stokes proof submitted in September.
- Google DeepMind's AlphaProof Nexus solved 9 of 353 open Erdős problems in May for a few hundred dollars, passing peer review by October.
OpenAI's GPT-6 Astra has helped produce a proof that could solve a 50-year-old problem in particle physics: how to place the full Standard Model on a computational lattice without generating unwanted ghost particles.
The result comes from Nathaniel Craig, a physicist at UC Santa Barbara who leads the particle theory group at UCSB and is a permanent member of its Kavli Institute for Theoretical Physics. Craig is currently a visiting scientist at OpenAI. He presented the approach on the OpenAI blog and in a preprint paper.
The story matters beyond physics. It extends a 2026 run of AI systems producing publishable mathematics — from an OpenAI disproof of an Erdős conjecture to a claimed solution of a Millennium Prize Problem — into theoretical physics, where AI is now contributing to proofs that human researchers have chased for decades.
Why the full Standard Model resists simulation?
The Standard Model is the theory describing all known particles and fundamental forces. Physicists can already simulate parts of it on a computer grid called a lattice. That technique has produced real breakthroughs. One of them is the calculation of the proton's mass from first principles.
But the full theory has never fit on a lattice. The obstacle is the weak force, which governs certain types of radioactive decay. The weak force treats left- and right-handed particles differently. On simple lattices, that asymmetry creates unwanted ghost particles — artifacts that wreck the simulation.
There is a known remedy. Special lattice particles called overlap fermions can filter the ghosts out. The catch has lasted more than 25 years: nobody could set the phase of the overlap fermions correctly.
What did Craig and GPT-6 Astra actually do?
Craig had a strategy for fixing the phase. His idea was to compare progressively finer lattices. What he lacked was a proof that the approach worked in all the cases it needed to cover.
That is where GPT-6 Astra came in. The model turned Craig's idea into a working construction. It then, in Craig's account, took the lead on the detailed proof itself.
The work did not proceed without human scrutiny. Lattice field theorists Maarten Golterman, Andreas Kronfeld, and Yigal Shamir reviewed the proof. They raised questions along the way. Craig says that process helped build confidence in the result.
He is careful about what the result means. He calls the work "much more a beginning than an end." A lot remains, he says, before realistic simulations of the full Standard Model become possible.
How does this fit the 2026 wave of AI mathematics?
The lattice proof did not arrive in isolation. It follows a run of AI advances in mathematics that started in January 2026.
The milestones, in order:
- In May, OpenAI's disproof of Erdős's unit distance conjecture became, per the source, the first AI proof worthy of mathematics' top journal.
- Days later, Google DeepMind's AlphaProof Nexus solved nine of 353 open Erdős problems — for a few hundred dollars.
- By October, AlphaProof Nexus's results had passed peer review.
- In September, OpenAI submitted a 166-page proof for the Navier-Stokes Millennium Prize Problem. The Clay Mathematics Institute said the problem had "apparently been settled."
The Navier-Stokes solution remains under review. Some mathematicians have questioned whether OpenAI fully understands its own proof. Across the field, mathematicians hold mixed feelings as AI keeps cracking previously unsolved problems.
What's at stake for physics?
Lattice methods are one of the few ways to extract precise numbers from quantum field theory without experiments. The proton mass calculation showed their power when the theory cooperates. The weak force's left-right asymmetry has kept the full Standard Model out of reach for half a century.
If Craig's phase construction survives further scrutiny, overlap fermions could finally become usable in full-theory simulations. That would open a path to first-principles calculations that today require particle accelerators — though Craig's own caution applies: the proof is a starting point, not a finished simulation pipeline.
Does AI co-authorship change how results get trusted?
The review process here is worth watching. Three named lattice field theorists — Golterman, Kronfeld, and Shamir — examined the proof and pressed on weak points. Their questions, in Craig's telling, strengthened confidence rather than undermined the claim.
That pattern echoes the broader 2026 debates. The Navier-Stokes case shows the tension: a 166-page AI-assisted proof, an encouraging but hedged statement from the Clay Mathematics Institute, and open doubt about whether the AI's creators understand what it produced.
The lattice result sits somewhere in between. A specialist proposed the strategy. A frontier model carried out the construction and the detailed proof. Domain experts checked it. Whether that division of labor becomes the standard workflow for theoretical physics is now a live question — and Craig's "much more a beginning than an end" framing suggests the field will find out case by case.
Original: openai.com
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- GPT-6 Astra Helps Physicist Propose Lattice Definition of Standard Model
- OpenAI says AI models refuted the Exponential Time Hypothesis
- OpenAI Claims GPT-5 Helped Close an Erdős Problem and Speed Immunology Breakthroughs
- OpenAI Claims Its AI Disproved a Core Assumption of Computer Science
- OpenAI says its models overturned a core computer science belief