The most arresting number in OpenAI’s mathematics announcement is not ten.
It is $2,000.
On August 1, OpenAI said an internal version of Astra, its next major model family, found new results for ten long-standing open problems in mathematics, quantum complexity, and theoretical computer science. The company estimates that the tokens used to find all ten solutions would cost roughly $2,000 at current GPT-5.6 Sol API rates.
That is a remarkable sentence, and it needs to be read with its units intact. OpenAI has not released Astra. There is no Astra API price, no Astra model picker in ChatGPT, Codex, or the API, and no launch date in the announcement. The $2,000 is not the cost of training Astra, or the price OpenAI paid to run a secret system, or a menu price a mathematician can pay today. It is a conversion: take the inference token count used in these discoveries and value it at Sol’s public API rates.
Even with those caveats, the number changes the story. Ten results spanning sphere packing, operator algebras, circuit lower bounds, quantum games, lattice cryptography, and Ramsey theory are not a benchmark. They are a research programme. Until this week, the public evidence for that programme was mostly an announcement and a claim. Now it has a price tag, papers, reasoning walkthroughs, and a public Lean repository.
The released model is GPT-5.6. Astra is the next one.
The timeline matters.
OpenAI released the GPT-5.6 family on July 9: Sol as the flagship, Terra as the balanced tier, and Luna as the low-cost tier. All three are available through ChatGPT, Codex, and the OpenAI API. Sol is the model whose rates OpenAI used for the Astra calculation. Astra remains internal, described by Noam Brown as OpenAI’s next major model family.
That makes the ten-proofs announcement a preview through evidence rather than a product launch. OpenAI is showing what its next system did before it tells developers what they can buy, how they will access it, or what it will cost.
The public GPT-5.6 rates are now unusually stark. As of July 30, Sol remains $5 per million input tokens and $30 per million output tokens. Terra was cut to $2 input and $12 output. Luna fell 80 percent to $0.20 input and $1.20 output per million tokens. Sol also gained a Fast mode, promising up to 2.5 times the speed at twice the standard price.
The cuts came three weeks after the family’s general release. OpenAI says the change followed serving improvements, including a 20 percent reduction in end-to-end serving cost from production-kernel work and more than 15 percent better token-generation efficiency. The company framed this as an engineering story: a stronger model was used inside a human-led process to optimize the systems that serve models, then some of the savings were passed through to customers.
The mathematics announcement makes it a scientific story as well. Sol is cheap enough to price the search. Astra is capable enough to make that search interesting.
What $2,000 does and does not buy
At Sol’s standard rates, $2,000 represents a serious inference budget. It is also small enough to be legible.
It is smaller than a research assistant’s monthly salary. It is less than many companies spend on a single long-haul conference trip. It is not the cost of a new university mathematics department, a cluster, or a new foundation model. Yet OpenAI says that much token-equivalent work produced ten new mathematical results, at least after selecting problems, running the search, and counting the tokens required for the successful solutions.
That phrase, “successful solutions,” is doing work. We do not yet know the full denominator: how many candidate problems Astra examined, how many paths failed, how much parallel exploration was discarded, or how much human and machine time went into choosing the ten problems. A price tag on the winning runs is not a total cost of scientific discovery.
Nor does it include the vast fixed cost behind the model. Training, data, infrastructure, safety work, evaluation, and the people who built the system are outside that $2,000 figure. So are the human efforts needed to turn arguments into manuscripts and the subsequent formalisation work.
OpenAI is admirably explicit about the last part. Its release says Astra generated the mathematical arguments; humans prepared the manuscripts with the same model; then the model formalised each argument in Lean. That is more informative than the usual blur of “AI-assisted research.” It gives the community a division of labour to interrogate. The machine made the arguments. People prepared the papers. The model produced the formal certificates. OpenAI takes responsibility for correctness.
That chain still deserves independent scrutiny. It is also far more concrete than asking everyone to trust a benchmark chart.
The proof package is the real release
The headline says an unreleased model solved ten problems. The more useful public object is the release package.
OpenAI published a paper, the model’s reasoning walkthroughs, and a GitHub repository containing Lean 4 formalizations. The repository has individual modules for the results, an All.lean build target, a stated Lean 4.32.0 toolchain, and Comparator challenge configurations. Anyone who has the environment can fetch the dependencies and run lake build All.
The ten results are not ten versions of the same trick. They include new upper bounds that reach the Cohn–Elkies threshold for high-dimensional sphere packing; stronger bounds for binary and spherical codes; a construction of a non-sofic group; a disproof of Connes’s rigidity conjecture; new lower bounds for computing the permanent; an exponential parallel-repetition theorem for quantum games; and a polynomial-factor hardness result for closest vector approximation. The package also claims results on Ehrhart’s volume conjecture, multicolour triangle Ramsey numbers, and two extremal graph-theory conjectures.
Formal verification does not settle every question that matters in mathematics. Lean can check that a specified formal statement follows from the listed axioms. It cannot, by itself, establish that the formal statement perfectly captures every informal phrase in a paper, or decide whether a technique is elegant, surprising, or fertile. Those remain jobs for mathematicians.
But verification changes the order of the conversation. The normal argument around a frontier-model announcement begins with whether to believe the result at all. Here, outsiders have artifacts they can inspect. The remaining questions become sharper: Are the formal statements faithful? Are the ideas new? Which parts of the pipeline were essential? What will specialists find when they read the proofs closely?
That is what OpenAI has bought with this release. It has made the claim expensive to ignore and possible to audit.
X saw two announcements, and it is right to keep them separate
The reaction on X split along exactly the line the company’s material draws.
One group saw a historic mathematics story. Sébastien Bubeck’s post led with non-sofic groups and pointed to the breadth of the work: von Neumann algebras, high-dimensional sphere packing, circuit complexity, and multicolour graph problems. Noam Brown called Astra a major step for scientific reasoning. The sense of scale is understandable. These are problems that were selected precisely because the central result had seen no progress for at least a decade, often much longer.
Another group asked the more durable questions. Outside mathematicians still need to verify that the papers say what the Lean statements say. The successful-token estimate needs a denominator. The release does not tell us whether Astra will arrive as GPT-6, GPT-5.7, or something separate, because OpenAI has not said. And nobody can order an Astra run for $2,000 today.
Those are not hostile objections. They are the difference between a result and a product.
GPT-5.6 is the product, and its economics are already public. Astra is the result, and its public face is a proof package rather than an endpoint. OpenAI has made the boundary unusually visible: a commercial model family whose falling prices set the yardstick; an internal successor whose work is being offered to mathematicians for inspection before its interface, price, or release date are known.
The $2,000 figure lands because it makes scientific search feel less like a distant moonshot and more like a bill someone might submit. It should not be mistaken for the full cost of the system that generated the work. It is still a startling measure of the marginal cost of asking a powerful research machine to keep trying.
The repository’s first instruction is not a vision statement. It is lake exe cache get. Then comes lake build All.