GPT-6.2 Astra produces the first known formal proof of a 2006 algebra theorem

Nov 18, 2026
Vals AI says Astra is the only model to solve the task on its Autoform-Hard benchmark, after a near give-up.

OpenAI's GPT-6.2 Astra is the only model to formalise a 2006 theorem by mathematician Agata Smoktunowicz on Vals AI's Autoform-Hard benchmark, the evaluation company reports. It is also the only formalisation of the theorem Vals AI says it has heard of.

Autoform-Hard has 56 research-level formalisation tasks. OpenAI's GPT-6.2 Sol and Anthropic's Claude Fable 5.2 each solve the same 48, and Astra solves a 49th: Smoktunowicz's result on the growth of graded algebras.

Working alone in a sandbox with Lean, Mathlib and a shell, Astra took 94 minutes, 146 agent turns and about 1,800 lines of Lean, at a cost of $21.58. About 47 minutes in, it submitted a partial proof with a gap, which the grader rejected, and filed an honest "unsolved" report. Because the benchmark does not let a model give up before spending $20, it was sent back to work, tried a new approach and had its full proof accepted 39 minutes later.

Vals AI says the proof passed clean-room grading and a hand review, uses only Lean's standard axioms, and matches the paper's theorem statement. After a shared opening, Astra's argument takes a different route from Smoktunowicz's published proof, according to Vals AI.

We hope you enjoyed this article

Consider subscribing to one of our newsletters like Daily AI Brief.

Also, consider following us on social media:

Free newsletter

Daily AI Brief

Daily report covering major AI developments and industry news, with both top stories and complete market updates

Whitepaper

Tensordyne Napier: What If One Rack Could Do the Work of Nine?

Tensordyne

This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.

Read more
Free, six days a week

Daily AI Brief: the AI news that matters, in your inbox.