GPT-6.2 Astra Formalizes a 2006 Theorem in Lean for $21.58

Vals AI says GPT-6.2 Astra has produced a Lean proof of a 2006 theorem by the mathematician Agata Smoktunowicz. In a blog post on 18 November 2026, Vals says the run took 94 minutes and cost $21.58 at Astra’s published price.

Our number: Astra solved 49 of the 56 tasks on Vals’ new Autoform-Hard benchmark, one more than Claude Fable 5.2 and GPT-6.2 Sol, which each solved 48, by our count from the leaderboard. That one task is the Smoktunowicz theorem.

Business Pill: The Cost of Checking

An explainer on who pays to check work, and why cheap verification changes what can be automated.

The key insight: Vals AI says GPT-6.2 Astra proved Smoktunowicz’s 2006 gap theorem in Lean in 94 minutes for $21.58, the one task that puts it at 49 of 56 on Autoform-Hard, ahead of Fable 5.2 and Sol at 48, by our count.

What Vals Says Astra Proved

The theorem says that no graded domain has Gelfand–Kirillov dimension strictly between 2 and 3. Vals says it is a result that, “as far as we know, had never been formalized”.

In plain terms, it is about how fast an algebra grows. Vals writes that the count grows “either no faster than about” n squared or at least as fast as n cubed, never in between.

Vals says the proof is about 1,800 lines of Lean. Its run log lists 1,794 lines of code and 146 agent turns.

GPT-6.2 Astra solved 49 of Autoform-Hard's 56 tasks, one more than Claude Fable 5.2 and GPT-6.2 Sol, by our co
GPT-6.2 Astra solved 49 of Autoform-Hard’s 56 tasks, one more than Claude Fable 5.2 and GPT-6.2 Sol, by our count from Vals AI’s leaderboard. Data: Vals AI, as reported, 18 November 2026.

What Was Posted on X

Vals AI announced the result on X on 18 November, calling it what it believes is the first Lean proof of Smoktunowicz’s Gap Theorem, on its new Autoform-Hard benchmark.

Andrew Curran quoted the passage from Vals’ post saying the published proof leans on three deep results not in Mathlib, and that Astra used none of them.

A Different Route From the Paper

Vals says “the argument takes a different route from the published one”. The published proof leans on three deep results that are not in Mathlib, Lean’s math library.

Vals writes: “Astra used none of them”. Instead, it says, Astra defined a quantity it calls “capacity” and built the rest of the proof on it.

Andrew Curran quoted that passage on X. Vals’ own post calls it “what we believe is the first Lean proof of Smoktunowicz’s Gap Theorem”.

Vals AI graphic of a number line from 0 to 5 with the gap between 2 and 3 marked, illustrating Smoktunowicz’s gap theorem
The gap the theorem rules out: no growth rate strictly between 2 and 3. Source: Vals AI, “GPT-6.2 Astra Formalizes Smoktunowicz’s Gap Theorem”, 18 November 2026.

How the Run Went

Vals says Astra worked alone in a sandbox with Lean, Mathlib and a shell, within limits of three hours and a $200 budget. It says “No human touched it”.

About 47 minutes in, Vals says, Astra submitted a partial proof with a gap, and the grader rejected it. Astra then filed an “unsolved” report.

The benchmark does not let a model give up until it has spent $20, and Vals says Astra had spent about $8. The harness sent it back to work, and the grader accepted the full proof 39 minutes after the report.

How Vals Says It Checked the Proof

Autoform-Hard accepts a proof only after clean-room grading. Vals says the statement must match the original exactly and the proof may use only Lean’s standard axioms; an independent kernel then replays it.

Vals says it also reviewed this proof by hand. It adds a caveat: “This is a formal proof of Smoktunowicz’s theorem, not a formalization of her proof”.

Where the Other Models Stand

The leaderboard page puts Astra first at 87.50%. Claude Fable 5.2 and GPT-6.2 Sol tie at 85.71%, and Gemini 4 Carbon is at 71.43%.

Astra’s listed cost per task is $15.24, about 34% below Fable 5.2’s $23.21, by our arithmetic. GPT-6.2 Sol lists $8.09.

The 56 tasks are ones that none of the ten frontier models in Meta FAIR’s original AutoformBench study solved, the methodology page says. AutoformBench has 341 Lean 4 tasks drawn from 24 graduate- and research-level textbooks.

Gemini 4 Neon, which Google launched the same evening, is not on the leaderboard as of 18 November. Earlier this week we covered OpenAI’s science blog on GPT-6 and Astra’s CelesteBench score.

The Structural Read

As we read Vals’ post, the grader did as much work as the model: it rejected a partial proof, refused an early give-up under the $20 rule, and only accepted a proof that a separate kernel replayed.

As we read the leaderboard, the top three models are separated by one task out of 56, so a single proof decides the ranking.

Vals AI, GPT-6.2 Astra Formalizes Smoktunowicz’s Gap Theorem

“Astra’s proof takes 94 minutes and about 1,800 lines of Lean, and after a shared opening it takes a different route from the published argument.”

Three Implications

FOR MODEL BUYERS Vals’ leaderboard lists Astra at $15.24 per task against $23.21 for Fable 5.2 and $8.09 for Sol on Autoform-Hard.

FOR MATHEMATICIANS Vals says it has annotated the full Lean proof line by line, with commentary on where each step sits relative to the published proofs.

FOR BENCHMARK READERS Vals says the 56 tasks are ones none of the ten frontier models in Meta FAIR’s original study solved.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The analysis describes a Verifier Moat: “built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”

As we read Vals’ post, Lean is that kind of verifier: its grader accepts or rejects a proof mechanically, and it rejected Astra’s partial attempt before accepting the full one. The analysis predates this result and does not discuss it.

What Is Not Established

The first-formalization claim is Vals’ own, hedged as “as far as we know”. We did not check the Lean code ourselves.

The post does not say whether outside mathematicians have reviewed the proof. Vals thanks OpenAI for permission to publish the proof and excerpts of the model’s reasoning; we did not contact OpenAI or Vals.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Four Intelligence Moats

A Business Engineer analysis of where durable advantage forms as AI models improve, including verifier-driven training.

Read the Map of AI →

The Bottom Line

Vals AI says GPT-6.2 Astra proved Smoktunowicz’s 2006 gap theorem in Lean in 94 minutes for $21.58. It is the one task that puts Astra at 49 of 56 on Autoform-Hard, ahead of Fable 5.2 and Sol at 48, by our count.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The proof, the run details, the checking process and the quotations are from Vals AI’s blog post of 18 November 2026, read in full that evening. Scores, costs per task and the benchmark’s design are from Vals’ Autoform-Hard leaderboard and methodology page, read the same evening. The 49-of-56 count and the 34% cost comparison are our arithmetic on those figures. We did not run the benchmark or check the Lean code, and did not contact Vals AI or OpenAI.

Sources: Vals AI: GPT-6.2 Astra Formalizes Smoktunowicz’s Gap Theorem · Vals AI: Autoform-Hard leaderboard · Vals AI on X · Andrew Curran on X

Scroll to Top
FourWeekMBA