The blog post went up quietly on August 1, with no launch event, no demo video, and no pricing page. In it, OpenAI said its next model, Astra, which has not been released to the public, had cracked 10 problems in mathematics and theoretical computer science that had resisted researchers for a decade or longer. The total compute bill for all 10 solutions: about $2,000, according to Startup Fortune and CNBC TV18.
In an industry where a single frontier training run can carry a price tag in the hundreds of millions of dollars, the figure reads like a typo. OpenAI’s point is subtler. The company says the problems were solved through reasoning at inference time rather than by spending more on training. Compute was aimed at the questions that needed it, instead of being poured into ever larger training runs.
The most prominent result, according to the company’s announcement, is the construction of the first known non-sofic group, a counterexample to a question that has stood since Mikhail Gromov introduced the notion of soficity in 1999. Other results span high-dimensional geometry, quantum information, and computational complexity. Roughly half are full resolutions of named conjectures; the rest tighten known bounds, which is progress of a different, lesser kind.
What sets the announcement apart from a typical AI research claim is the packaging. Each proof was formalized in Lean 4, a proof assistant, and the certificates were published on GitHub. Anyone with the tools can rerun the checker rather than take OpenAI’s word for it. The company also released a manuscript of roughly 250 pages describing the arguments.
The $2,000 figure needs a footnote, analysts said. It covers the cost of inference tokens at current API rates, the compute spent searching for the proofs. It does not include the training of Astra itself, the human mathematicians who selected the problems and shaped the arguments, or the years of verification work that may follow. It is a statement about the price of reasoning, not the price of the research.
Even so, the number has landed in the middle of a live argument. For months, skeptics of the so-called scaling law have argued that the industry is running out of returns on bigger models and more data. Astra, if its results hold up, offers a different lesson: that a model can get smart enough to spend a modest amount of compute where it counts, turning inference into a tool for discovery rather than a bill to be paid.
The arithmetic behind the announcement has been picked apart since. At current API rates, the company’s estimate works out to roughly $200 per solved problem, and OpenAI has said the figure reflects the tokens spent finding the solutions rather than the full cost of the research program. Even discounted, the comparison to conventional computational mathematics is striking. Labs have historically rented supercomputer time by the month to make progress on problems of this difficulty; Astra’s entire haul fits inside the budget of a mid-range laptop purchase.
The timing is also notable. Astra remains unreleased, with no confirmed launch date, and the company has said the model must pass a US government security review before any public rollout. That puts a hard regulatory checkpoint between the math results and the product, and it means the work will circulate in academic channels for some time before it reaches paying customers.
For researchers, the economics are the headline. A compute budget that any well-funded university group could authorize has produced results that previously required a research career. OpenAI separately said it is giving 100,000 academic researchers free access to its current frontier models through 2027, a program that runs alongside the Astra work and deepens the company’s ties to the scientific community.
The broader shift, analysts said, is the pairing of long-horizon reasoning with formal verification. Let a model think for hours on a single problem, then check the output mechanically. That combination travels beyond mathematics, into software correctness and other domains where a proof can be expressed formally, and it may be the part of the announcement with the longest reach.
For the AI industry, the stakes are commercial as much as scientific. If frontier models can do original mathematics cheaply, the value of reasoning capability grows relative to raw training scale, a shift that would touch how labs spend, how chips get used, and how the next generation of models is priced. Test-time compute, long a footnote in the cost debate, moves to the center.
The caveats are real. Machine-checked formal correctness is not the same as community acceptance of the underlying ideas, and mathematicians will spend years digesting the constructions. No Millennium Prize problem was touched. But the direction is clear: the cost of discovery has collapsed, and the ability to prove what a model found has become the binding constraint.
What to watch now is verification. If independent researchers confirm the non-sofic construction and the other results, the $2,000 line will enter the debate over AI’s limits as one of the cheapest experiments in modern mathematics. If parts unravel, it becomes a cautionary tale about announcement over evidence. Either way, OpenAI has succeeded in putting a price tag on a question the industry has been arguing about in the abstract.
Either way, the price of a discovery has just become a number people can argue over.
