A 249-page paper arrived with a name attached
Sebastien Bubeck, a researcher at OpenAI, spent Saturday telling the internet that nonsofic groups exist. He described the statement as one of many results proved by Astra, which he called the company's next major model, and said OpenAI was releasing ten such proofs complete with Lean certificates and reasoning walkthroughs for each of them. Attached to the announcement were a 249-page manuscript, a second document of reasoning walkthroughs, and a public repository of machine-checkable certificates. The coverage that followed settled quickly on two numbers: ten problems, and about 2,000 dollars of compute.
The manuscript itself never uses the word Astra. It opens by presenting a collection of results obtained by an internal OpenAI model, and then spends 249 pages on sphere packing, von Neumann algebras and circuit complexity without naming the thing that produced them. The name lives in the announcement. The mathematics lives in the paper. Keeping those two surfaces separate is the whole job for anyone reading this as a buyer rather than as a spectator.
What actually shipped is genuinely checkable
The ten results are not decorative. The first determines the exact asymptotic strength of the Cohn-Elkies linear program, raising the general high-dimensional sphere-packing exponent from the classical Kabatianskii-Levenshtein value of 0.59905576 to about 0.6044, which the paper states is the first improvement to that general bound since 1978. The third constructs an explicit non-sofic group, resolving whether every countable group admits finite permutation approximations. The fourth disproves Connes's rigidity conjecture by building infinitely many pairwise nonisomorphic property-(T) groups sharing one group von Neumann algebra. The rest cover binary and spherical codes, lower bounds for computing the permanent, quantum parallel repetition, the closest vector problem, Ehrhart's volume conjecture, a superexponential lower bound for multicolour Ramsey numbers, and counterexamples to two conjectures in extremal graph theory, three of them drawn from Paul Erdos's catalogue.
Each one carries a Lean 4 formalisation in a public repository, built against mathlib on Lean 4.32.0 and released under an Apache-2.0 licence. That is the part of this release that deserves unqualified credit. A machine-checkable certificate means a stranger with a laptop can confirm the argument holds without trusting OpenAI, without trusting the model, and without waiting for peer review. Most capability claims an owner has been handed this year arrived as a benchmark score published by the party selling the model. These arrived as objects that answer to neither.
The acknowledgement names a model you can already buy
At the end of the fourth chapter, after the thanks to Sorin Popa and others for comments and readings, sits one sentence that no coverage picked up. During the preparation of this manuscript, OpenAI writes, it learned of independent and concurrent work by Shuoxing Zhou also establishing a counterexample to Connes's rigidity conjecture, developed in part with the assistance of GPT-5.6 Sol. Sol is not the unreleased model. Sol is the model OpenAI sells today, with a published price list.
Read plainly, one of the ten headline results was reached at the same time, by someone outside the lab, working with generally available software. That does not diminish the other nine, and OpenAI disclosed it in its own document rather than leaving it to be discovered, which is to the company's credit. But it is the single most decision-relevant sentence in the entire release for anyone deciding what to license, and it sits in an acknowledgements paragraph on the far side of 200 pages of operator algebra rather than in the announcement.
A price tag borrowed from a different product
The 2,000 dollar figure has done more work in the last day than any other number in the release, and it is not a price. It is a token count valued at the published API rates of GPT-5.6 Sol. Astra has no published rates, because Astra is not for sale. So the figure answers a hypothetical question, namely what those tokens would have cost had they been billed as Sol tokens, and it is being read as an answer to a different question, namely what frontier mathematical research now costs.
What the number does not carry is as important as what it does. It does not disclose how many attempts preceded the ten that worked, how much human direction shaped them, how long any of it took in wall-clock terms, or what Astra will cost when and if it ships. OpenAI has given it no release date and describes it only as the next major model family. An owner who files 2,000 dollars away as the going rate for a solved conjecture has recorded a number that was never a bill for a product that cannot yet be bought.
Ask what the shipping model does
The gap between the model in the demonstration and the model on the price list is the most reliably manipulated variable in AI procurement, and it is usually invisible because vendors are not obliged to document it. This release is unusual precisely because the evidence about the shipping model is inside the vendor's own manuscript. The correct response is not scepticism about the mathematics, which is formalised and checkable, but discipline about the inference. A verified proof is a claim about the proof. It says nothing about which model produced it, on which attempt, with how much help, or at what price.
Three questions turn this into procurement practice. Which model version generated the result you are showing me, and is that version generally available today. If it is not, what does the generally available version do on the same task. And what are the published rates for the version I would actually be running. A vendor that can answer all three is selling a product. A vendor that can only answer the first is showing you a research preview, however impressive the artefact attached to it.
Read next: Microsoft Trained Its Sellers to Call Claude Slower | Google Scrapped a Near-Finished AI Model
