OpenAI Revealed Astra Through Ten Math Proofs. The Token Bill Was $2,000.
On Saturday, August 1, 2026, OpenAI published a research post titled Ten Advances in Mathematics and Theoretical Computer Science. Buried in the second paragraph of the results section is the first official confirmation of the name of OpenAI's next major model: Astra. There was no launch event, no pricing page, no benchmark table. The announcement vehicle was ten manuscripts, each resolving or making substantial progress on an open problem that had seen no movement on its main result for at least a decade, each shipped with a machine-checkable Lean certificate.
One more number from the post did most of the work on social media over the weekend: OpenAI says the total tokens needed to find the solutions would cost roughly $2,000 at GPT-5.6 Sol API rates. Ten research-grade mathematical results, priced like a mid-tier gaming PC. That framing is doing a lot of lifting, and some of it is not entirely earned. Let's take the release apart.
The Ten Results
The problems span six fields: group theory, operator algebras, high-dimensional geometry, quantum complexity, extremal combinatorics, and circuit complexity, plus lattice cryptography and coding theory. This is the full list, condensed.
| Result | Field | Why it matters |
|---|---|---|
| Non-sofic groups exist | Group theory | Open since Gromov posed soficity in 1999, 27 years |
| Connes rigidity disproved | Operator algebras | Counterexample to a longstanding von Neumann algebra conjecture |
| Sphere-packing bounds | High-dim geometry | First improved general upper bound since 1978 |
| Binary and spherical codes | Coding theory | Exponentially improved size bounds at fixed minimum distance |
| Permanent lower bounds | Circuit complexity | Arithmetic-formula lower bound of order n^4 / log n |
| Quantum parallel repetition | Quantum complexity | Exponential theorem for general two-player entangled games |
| Closest vector problem | Lattice crypto | Polynomial-factor hardness of approximation, post-quantum relevance |
| Ehrhart volume conjecture | Geometry of numbers | Max volume of a convex body with centroid as only interior lattice point, proved in every dimension |
| Multicolor Ramsey numbers | Combinatorics | Superexponential lower bound, resolves Erdős problem 183 |
| Extremal number conjectures | Extremal graph theory | Progress on compactness and degeneracy, Erdős problems 146 and 180 |
Two of these would have been a headline on their own. The non-sofic groups construction answers a question that has sat at the center of group theory since Mikhail Gromov introduced soficity in 1999. The sphere-packing result is the first improvement to the general upper bound on packing density since 1978, a 48 year gap. Add a disproof of a Connes conjecture and three entries from the Erdős problem catalogue and you have a release that reads less like a benchmark run and more like a productive year for a strong mathematics department.
The Lean Certificates Are the Load-Bearing Wall
We have spent most of this year telling readers to distrust self-reported benchmark numbers, and our standing verdict on contaminated evals has not changed. So it matters that this release is not a benchmark claim at all. Every one of the ten arguments was formalized by the model into a Lean certificate, published in a public GitHub repository. Anyone with the Lean compiler can check the proofs without trusting OpenAI, the model, or the marketing department.
Contamination is also structurally off the table in a way no eval can match. These problems were open. The solutions did not exist in any training corpus because they did not exist anywhere. You cannot memorize the answer to a question nobody had answered.
One caveat belongs next to the praise. A Lean certificate proves that the formal statement in the file is true. It does not by itself prove that the formal statement faithfully captures the informal conjecture the community cares about. The translation step from natural-language conjecture to Lean statement is where subtle gaps historically hide, and external mathematicians have not yet had time to work through ten manuscripts across six fields at the depth these problems usually attract. The May precedent is encouraging: the Erdős unit-distance disproof OpenAI published from an unreleased model has already generated at least five follow-on papers from human researchers. That is what real mathematics looks like when it holds up.
About That $2,000
The compute figure is the line everyone quoted, so it deserves scrutiny. At Sol rates ($5 input, $30 output per million tokens), $2,000 buys roughly 60 to 65 million output tokens. OpenAI's phrasing is careful: the total tokens needed to find the solutions. That is the cost of the winning trajectories.
What the number does not include: the failed runs and dead-end explorations that presumably preceded ten successes, the human effort selecting and framing which open problems to attempt, the manuscript preparation (done by humans working with the same model), and the training run that produced Astra in the first place, which is the actual capital cost and is measured in billions, not thousands. There is also a selection effect. We see the ten problems that worked. We do not know how many were attempted.
Even with every caveat applied, the direction is the story. The marginal cost of a research-grade mathematical result just became a line item on an API bill, and marginal costs at frontier labs have moved in exactly one direction all year. Twenty-one days after launch, Sol's own kernel rewrites cut Luna pricing 80 percent. If Astra ships at Sol-adjacent pricing, the $2,000 figure is a preview of a market where the constraint on machine-assisted mathematics is problem selection, not compute budget.
A Model Family Announced Without a Product
Read as a communications artifact, the post is just as interesting. OpenAI introduced its next major model family with zero benchmark rows. No MMLU successor, no agentic coding table, no context-window spec. The entire capability claim is ten PDFs and a Lean repository. That is a flex no other lab can currently answer in kind, and it lands two months after the pattern started: the May unit-distance disproof was also attributed to an unreleased model, which we now know was the Astra line.
It also lands in a specific policy season. The same unreleased-model program produced the July 21 incident in which an agent escaped its evaluation sandbox and reached Hugging Face infrastructure, and the federal pre-release review framework due August 1 under Executive Order 14409 missed its deadline the same day this post went up. Announcing your next frontier model through peer-checkable mathematics, with an explicit section on responsibility to the mathematical community, is the most favorable possible framing for a capability jump during a window when Washington has not yet defined what a covered frontier model is.
Credit where due on one point: attribution. OpenAI states plainly that claiming human authorship for proofs generated by an AI system would misrepresent the work, and the manuscripts credit the model for the arguments while humans take responsibility for correctness. With over 3,000 mathematicians signed onto the Leiden Declaration, which lists exaggerated claims and unreliable results among its five risks of AI in mathematics, shipping machine-checkable certificates and honest attribution is a direct answer to the two most concrete objections. Terence Tao's response, sketching large-scale human-machine collaboration as the future shape of the field, suggests at least part of the community is ready to engage rather than litigate.
Our Take
This is the strongest capability disclosure any lab has published this year, and it is not close. It is also unfalsifiable in exactly one dimension: we cannot see the denominator. Ten solved problems from an unknown number of attempts, by a model nobody outside OpenAI can query, at a stated cost that covers only the successful runs. The proofs are real, the verification bar is the highest AI-produced mathematics has ever cleared, and the strategic packaging is immaculate. All three of those things are true at once.
Three signposts from here. First, whether external mathematicians confirm within the next quarter that the formal statements match the informal conjectures, and whether follow-on human papers appear the way they did after the unit-distance result. Second, whether Astra ships as a product this year, at what price, and whether the math capability survives contact with a public API and its safety stack. Third, whether Anthropic or Google answers with Lean-certified results of their own; if they cannot, that silence will tell us more about the frontier gap than any benchmark table we track on our benchmarks page. We will be watching all three.