Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

GPT-6 Astra and a candidate lattice proof

On November 15, 2026 OpenAI's science blog ran Nathaniel Craig's candidate lattice Standard Model, built with GPT-6 Astra. The proof is still a candidate.

RA
· Founder, AIToolRanked
A dark field of amber and blue glowing points joined by thin lines in a receding lattice
On this page

What OpenAI published on November 15

On November 15, 2026, OpenAI posted A step toward defining the Standard Model on a lattice ↗ as the first guest post on its science blog ↗. The author is Nathaniel Craig. He leads the particle theory group at the University of California, Santa Barbara, is a permanent member of the Kavli Institute for Theoretical Physics, and is a visiting scientist at OpenAI. The science hub lists that post, and a note introducing the blog, both dated November 15.

The picture above is ours. We did not check the proof. We did not run the model.

Craig's line in the post: "GPT-6 Astra was essential to turning the refinement idea into a working construction." He also writes that Astra "took the lead in developing the detailed proof." He says lattice theorists Maarten Golterman, Andreas Kronfeld, and Yigal Shamir "spent time looking over Astra's proof and asking probing questions." The result, in his words, is "a candidate nonperturbative lattice definition for the Standard Model." He names two open problems: whether the construction satisfies reflection positivity at finite lattice spacing, and a definition of the interacting theory in continuous spacetime. He calls it "much more a beginning than an end."

What the October preprint actually locks down

The paper is arXiv:2610.09093 ↗, "A gauge-invariant measure for lattice chiral fermions," version 1 submitted October 6, 2026 at 20:47 UTC. The abstract calls it "a candidate nonperturbative lattice definition of the Standard Model." Craig's OpenAI post describes the same October preprint.

The acknowledgements say the proof is Astra-led after rounds with Craig and outside experts, and that the manuscript mixes Craig's sections with rewrites of Astra's proof. He writes that he and colleagues proposed tests, including Lean verification, and that Astra ran them. A supplement says all sixteen Lean 4 modules compiled clean, and a kernel replay accepted the proofs on Lean's standard axioms with no placeholders. Those checks cover the formal statements under their assumptions. AI agents, not the physicists, reviewed the match to the mathematical companion. Craig writes that Lean is not the last word: a conclusion can follow from its definitions and still be the wrong physics.

The last section says the formulation is one of principle rather than practice, and again lists reflection positivity and a definition of the interacting continuum theory as open. That matches the blog.

The acknowledgements split the three theorists more finely than the blog sentence. Golterman and Shamir are thanked for questions across the whole project. Kronfeld is thanked for expertise and comments on the manuscript. A claim that Edward Witten built on this paper is not in the post or the preprint we read. It is not in this note.

What Astra costs, and which GPT-6 this is not

On OpenAI's GPT-6 Astra model page ↗, fetched November 16, 2026, text tokens are $10 input, $1 cached input, and $50 output per million. A prompt with more than 272K input tokens is priced at 2× the input and cache rates and 1.5× output for the whole request. That is the API list on that page. It is not a ChatGPT receipt.

This model is not GPT-6 Cyber, the gated cyber model from November 5. Deepnote's Agent Initiative Benchmark is a vendor Slack test that printed Astra at 8.97 of 10. Deepnote's writeup says that task set does not separate Astra from Sol. Fable 5.2, in our November 10 note, is $8 and $40 per million tokens. A lattice measure and a Slack agent test are different jobs. AlphaProof Code is the other Lean story here: DeepMind agents on Signal's compiled code, not a fermion measure.

What people are saying

  • "The word "candidate" matters. It is a starting point for other physicists to check, not a solved problem." — Akshay Anil Dewalwar ↗ (@Adewalwar), November 16

  • "Another huge AI assisted physics breakthrough" — Dr Singularity ↗ (@Dr_Singularity), November 16

  • "I want to throw a stuck problem at this model." — Claire ↗ (@Clairef9vu), November 16

Reddit was not reachable for this story.

Our take

Read "candidate" before you read "breakthrough." Craig, the abstract, and the paper's last section all keep the word. Reflection positivity and the continuum limit are named as open. Sixteen Lean modules and a clean kernel replay check that the formal statements follow from their assumptions. Craig says AI agents reviewed the match to the companion, and he says that review is not a physics sign-off. Golterman and Shamir worked the project. Kronfeld commented on the manuscript. The blog says all three looked over the proof. That is not a community verdict.

Call Astra on a stuck formal argument only when the bill is acceptable and a checker exists. Past 272K input tokens the list price jumps. A lattice post is not a reason to point a coding agent at production math and ship the file. Craig's check is Lean, plus people who already know the problem. METR's November 5 report measures the pace of Anthropic's own research. It does not score this construction.

Who should care. People who work on chiral lattice gauge theory, and anyone buying a research model who wants to see what "the model led the proof" means when the author still says candidate. Who should not restack a product on this post. Teams pricing a coding agent. Astra at $50 per million output tokens is a different purchase from Fable 5.2 at the $8 and $40 in our November 10 note, and from the Slack-test gap in the Deepnote writeup. Dr Singularity's "breakthrough" is his word. Craig's word is candidate.

More from the news desk · About Rai Ansar