OpenAI Astra: 10 Math Proofs, Verified in Lean (2026)

OpenAI announced Astra on August 1, 2026 by publishing ten new results on mathematics and theoretical computer science problems that had been open for at least a decade. An internal version of the unreleased model generated the arguments, then formalized each one as a machine-checkable Lean certificate anyone can download and verify.

TL;DR

• OpenAI published ten results on August 1, 2026, produced by an internal version of Astra, which it calls its "next major model."¹
• Headline result: a construction establishing the existence of non-sofic groups — a central open question in group theory since Gromov posed it in 1999.¹²
• Three of the ten resolve numbered problems from the Erdős catalogue: 183, 146, and 180.¹
• Every result ships with a Lean 4 certificate in a public Apache-2.0 repository.³
• OpenAI says the tokens "would cost roughly $2,000 at Sol API rates."¹ That figure covers the attempts that worked — OpenAI researcher Noam Brown confirmed other major problems were attempted without success.⁴
• Astra is reported to be a multi-agent system trained for long-horizon tasks running hours or days. It is unreleased, with no ship date.⁵

What You'll Learn

• What OpenAI actually announced, and which ten problems were solved
• How Astra's reported multi-agent architecture differs from a single long context window
• Why the Lean certificates — not the math — are the real story for agent builders
• What the $2,000 figure does and does not cover, and why it is a cost per success
• Which parts of the announcement remain unverified

What OpenAI announced

Astra is OpenAI's next major model family, sitting alongside the existing Sol, Terra, and Luna lines.⁵ It has not been released, has no announced ship date, and OpenAI has not decided whether it ships as GPT-6 or as a variant in the GPT-5 line.⁵

The August 1 publication is the first time OpenAI has publicly confirmed the Astra name.⁵

OpenAI's framing is specific. These are problems that "have been open and have seen no progress on the main result for at least a decade, and in most cases much longer."¹ They span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics.¹

The ten results

Table: The ten results as stated by OpenAI. Source: OpenAI, "Ten advances in mathematics and theoretical computer science," August 1, 2026.

The non-sofic group construction is the one drawing most attention. Soficity was introduced by Mikhail Gromov in 1999, and named by Benjamin Weiss in 2000 after the Hebrew word for finite.² Gromov's paper posed the question that stood ever since: is every group sofic?²

Astra is a multi-agent system, not a bigger context window

This is the part that matters if you build agents.

According to reporting by The Information, Astra is a multi-agent system trained specifically for long-horizon tasks — a root agent creates subagents, distributes portions of a problem, waits for results, and synthesizes a final answer.⁵ OpenAI has not published an architecture paper, so treat the internals as reported rather than confirmed.

What OpenAI has confirmed is the shape of the workload: systems that keep working on one objective for hours or days.⁵

That distinction matters. The dominant failure mode in long-running agents is not a weak base model — it is compounding error, where a small drift early in a run gets amplified across every subsequent step because nothing in the loop can tell the agent it went wrong.

Multi-agent setups do not automatically fix this, and there is now hard data on when they make it worse. A study from Google Research, Google DeepMind, and academic collaborators — detailed by Google Research on January 28, 2026 — evaluated 180 agent configurations across five architectures and three model families.⁶

The results cut both ways:

Table: Selected findings on multi-agent scaling. Source: Google Research, "Towards a science of scaling agent systems," January 28, 2026.

Note which architecture contains errors. Independent agents working in parallel without communicating amplified mistakes 17.2×; adding a central orchestrator cut that to 4.4×, because the orchestrator acts as a validation bottleneck that catches errors before they propagate.⁶

That is the same shape as Astra's reported design — a root agent that delegates, waits, and synthesizes. If the reporting is accurate, OpenAI picked the architecture the evidence favours for error containment.

Astra's results are evidence that the pattern can work at research scale. They are not evidence that it works by default, and mathematical proof search is unusually parallelizable — you can explore many independent attack routes at once. Do not assume the same gains transfer to a tightly sequential workflow.

If you are wiring up delegation between agents, our walkthrough of agent-to-agent handoff in TypeScript covers the mechanics.

The Lean certificates are the real story

Here is what most of the coverage skipped.

Every one of the ten results ships with a Lean 4 formalization in openai/ten-proofs, an Apache-2.0 repository built against Lean 4.32.0 and mathlib.³ Two commands — lake exe cache get and lake build All — check all ten.³

That is a deterministic oracle bolted onto a stochastic generator. Anyone who distrusts OpenAI entirely can still confirm that the formal statements follow from the axioms, without taking a single claim on faith.

This closes the exact failure mode mathematicians had been warning about: arguments that read as plausible, are wrong, and are expensive for humans to audit.

But be precise about what a Lean certificate proves. It proves the formal statement follows from the axioms. It does not prove the formal statement faithfully captures the informal theorem it claims to encode. If the formalization quietly weakens a hypothesis, Lean will happily verify a weaker result. Auditing the statements against the informal claims is separate work, and it is the fastest independent signal available on this release.

Terence Tao gave this shape a name on December 15, 2025: artificial general cleverness rather than intelligence.⁷

His definition is worth quoting in full, because it describes this release almost exactly:

"By 'general cleverness', I mean the ability to solve broad classes of complex problems via somewhat ad hoc means. These means may be stochastic or the result of brute force computation; they may be ungrounded or fallible… And yet, they can have a non-trivial success rate at achieving an increasingly wide spectrum of tasks, particularly when coupled with stringent verification procedures to filter out incorrect or unpromising approaches, at scales beyond what individual humans could achieve."⁷

Tao's recommended framing is to treat these systems "primarily as a stochastic generator of sometimes clever — and often useful — thoughts and outputs."⁷

That is a design constraint, not a philosophy. A stochastic generator becomes trustworthy in proportion to the verifier attached to it — which means the ceiling on your agent's autonomy is set by how stringent your checking is, not by how good your model is.

It generalizes far past mathematics. A type checker, a passing test suite, a schema validator, a simulator, a reconciliation query: each is a cheap deterministic oracle that lets an unreliable generator run longer without the output degrading into plausible nonsense.

This is the same pattern showing up in self-verifying agents in chip design, where agents call deterministic physics engines to validate their own decisions.

What the $2,000 actually buys

OpenAI's wording is careful: the tokens needed "would cost roughly $2,000 at Sol API rates."¹ Brown put it at "under $2,000 at Sol API prices."⁸

Read that closely. Sol — not Astra. Astra is unreleased and unpriced, so this is a counterfactual: what the token volume would have cost billed at a different model's rates. No such amount changed hands, and the figure tells you nothing about what Astra itself costs to run.

Divided evenly, $2,000 across ten results is about $200 per solved problem.

Then comes the caveat that changes the arithmetic. Brown, in the same thread, wrote:

"And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet). But also, we didn't spend a lot on each problem. It's possible to push test-time compute much further."⁴

So the $2,000 is a cost per success, published without a success rate. The failed attempts are excluded. There are seven Millennium Prize Problems, selected by the Clay Mathematics Institute in 2000 at $1 million each; only the Poincaré conjecture has been solved.⁹

Credit where it is due — Brown volunteered the failures unprompted and in public. But a numerator without a denominator is not a cost per result.

Worth noting: even if the true rate were one success per hundred attempts, the cost per major open problem would land near $20,000. For results of this kind, that is not an expensive number. The economics survive the correction; the precision does not.

This is the same trap covered in AI agent token costs in 2026: per-token prices keep falling while cost-per-completed-task climbs, because the failures never show up in the headline figure. If you are budgeting a long-horizon agent, the number you need is cost per attempt multiplied by attempts per success.

What's still unverified

Three things are true at once, and most of the argument online consists of asserting one as if it settled the others.

The proofs are very likely correct. Lean certificates are strong evidence, subject to the statement-matching caveat above.

The significance is unestablished. That requires the relevant specialists to digest the results, which takes months.

The economics are unknown. No success rate has been published, and no list of failed attempts.

External reproduction is currently impossible while Astra stays internal. And OpenAI is explicit about attribution: it helped prepare the manuscripts and formalize the proofs and "take[s] responsibility for their correctness, while the mathematical arguments themselves were generated by our system," citing the Leiden Declaration on AI and Mathematics on how credit should be assigned.¹

One further limit on the companion reasoning walkthroughs: they narrate the model's thinking on problems it solved.¹ There is no walkthrough for the problems it failed. What is published is the search behaviour of the winners.

The bottom line

The number that travelled was $2,000. The number that matters is zero — as in, zero trust required to check the results.

Ten machine-checkable certificates in a public repo is a materially different artefact from a benchmark score, and it is the part of this announcement that generalizes. Every agent team shipping long-horizon systems should be asking the same question OpenAI answered here: what is the cheapest deterministic oracle that can tell my agent it is wrong?

Get that right, and you can let the loop run longer. Get it wrong, and more autonomy just produces more confident slop.