Start with the sentence almost nobody quoted.

Buried in OpenAI’s August 1 announcement, right after the claim that an internal version of Astra produced ten new results in mathematics and theoretical computer science: “The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.”

Two thousand dollars. For work on problems that had been open for a decade or more, including the first explicit construction of a non-sofic group and a counterexample to Connes’s rigidity conjecture from 1980.

The money stopped being the filter

If you run technology at a mid-sized company, that number should reorganize your thinking more than the proofs do.

Two thousand dollars does not trigger a procurement process. It does not need a business case, a steering committee, or a line in next year’s budget. It is roughly what a team spends on a single conference. Whatever it is that stands between your company and this class of capability, cost is not it anymore.

What stands between you and it is that most companies cannot name the problem they would point it at.

Aaron Levie, Box’s CEO, caught this the same day in a post on X that is sharper than most of the coverage. Capability in deep domains, he argues, is about to go vertical while everyday use barely moves. Then the part worth pinning to a wall: this “will often lead to a capability overhang, though, as many of these performance gains need to be applied to data sets and workflows in an applied way for that area of work.”

Capability overhang is the right phrase. The model is now above the problem. The missing piece is the machinery that connects one to the other, and that machinery is organizational, not technical.

What the announcement actually claims

I want to separate the checkable from the promotional, because both are present.

Checkable: the manuscript runs 249 pages and credits the results to “an internal OpenAI model.” The openai/ten-proofs repository carries one .lean file per result, builds with Lean 4.32.0, mathlib and Lake, and ships under Apache 2.0. That is the real break from every previous AI-does-math moment. Lean’s kernel does not care who wrote the proof or how good the blog post was. You clone the repo, run lake build All, and reach your own verdict. OpenAI also published 62 pages narrating how the model got there.

The spread matters as much as the depth. The ten results reach across high-dimensional sphere packing, binary and spherical codes, group theory, operator algebras, arithmetic circuit complexity, quantum parallel repetition, lattice cryptography and extremal graph theory, and two of them close three separate entries in Paul Erdős’s problem catalogue, including problem 183 on multicolor triangle Ramsey numbers. This was not one lucky domain where the training data happened to be dense. It was eight.

Now the careful reading of the money. OpenAI does not say “we spent $2,000.” It says the tokens needed to find the solutions would cost about that at Sol rates, which is a retail price applied in hindsight rather than an invoice. The same paragraph continues: the arguments “were then prepared into manuscripts by humans with the same model,” and afterward “the model formalized each argument in a Lean certificate.” None of that downstream work sits inside the $2,000. Neither does whatever compute went into attempts that led nowhere.

That distinction does not survive retelling. Linas Beliunas’s LinkedIn write-up, one of the widely shared summaries, compresses it to “total inference cost: approximately $2,000.” That is a firmer claim than OpenAI made, and it is one short step from there to “OpenAI spent $2,000.” If you are taking this number into a board meeting, take the original phrasing with you.

For scale, GPT-5.6 Sol lists at $5 per million input tokens and $30 per million output tokens according to OpenRouter’s price tracker. If every token were output, $2,000 buys around 67 million of them. The actual mix is not published, so treat that as my arithmetic on a public rate rather than an OpenAI figure. It fixes the order of magnitude and nothing more: tens of millions of tokens.

And the caveat that matters most commercially: Astra is not for sale. OpenAI calls it the next major model, with no ship date. What you can buy today is not Astra. What you can build today is the organization that will know what to ask it.

The verification part is worth copying before the model is

Two months before this announcement, in June, the mathematical community published the Leiden Declaration on Artificial Intelligence and Mathematics, endorsed by the International Mathematical Union. One of the threats it lists reads like it was drafted for August 1: “Proper evaluation is endangered if results are communicated through informal channels such as press releases or blog posts, often without any research paper or other disclosure of information necessary for scientific evaluation.”

OpenAI acknowledges the tension directly, and takes a position I think is right: claiming human authorship for a proof generated entirely by an AI system, the announcement says, “would misrepresent both the system’s contribution and the nature of genuine human intellectual work.” The company takes responsibility for correctness while crediting the arguments to the system.

The underlying Leiden point still holds, and it generalizes far past mathematics. A result without an independent verification path is not a result, it is a claim. The reason this release survives scrutiny better than the AI math announcements before it has nothing to do with the model being smarter. It has to do with a compiler that returns yes or no.

I have been programming for 36 years, since 1990, and I have never had a tool that handed me its answer along with a certificate I could check without trusting the source. That property, more than the ten proofs, is the one I want companies to steal. Most enterprise AI work today produces output that only its author can evaluate. Somebody reads it, it looks plausible, it ships. There is no lake build for a rewritten quarterly forecast.

You need a mechanism, not a model

So what does IQ Source do about this?

We have already argued that price per token lies and you should measure cost per job, and that the harness is the moat once the model is a commodity. Astra pushes that thesis to its limit. When the compute needed to attack a genuinely hard problem costs less than a laptop, price is no longer a filter. The only filter left is judgment about where to aim.

That judgment does not get produced in a workshop. It comes from a written inventory of how the company actually operates, where time and money leak, and how checkable each result would be if you automated it. Then a ranking of those candidates that nobody can override on preference alone.

If you want a quick screen for whether an internal problem is a candidate for this class of compute, three conditions do most of the work. One: something other than the person who asked for the answer can confirm the answer is right, whether that is a database query, a reconciliation, or a test suite. Two: the problem has been stuck for a reason other than nobody having found the hours, because if effort is the only missing input then you have a staffing question, not a capability question. Three: the cost of being wrong is bounded and surfaces before it reaches a customer. The ten Astra results satisfy all three, which is precisely why they worked as a demonstration. The first condition is the one most enterprise AI work quietly skips, and skipping it is what leaves a project impossible to defend to anyone who did not write it.

That is exactly what AI Maestro is built to produce. Two months of consulting, education and training alongside your team, not an installation. It outputs two things: a Process Reality Map, which documents how the business runs in practice rather than how the handbook describes it, and an AI Opportunity Score, which ranks the candidates by impact and by how verifiable the result would be. The program closes at a Go/No-Go gate. Design and implementation are a separate, conditional stage, because sometimes the honest answer is “not yet.”

A CTO holding that map is in a different position from everyone else. When Astra ships, or the equivalent from Anthropic or Google, they can say within the hour which problem deserves that compute and how they will prove it worked. Everyone else spends the next two quarters in a committee deciding where to start, while the capability sits there, already affordable.

The $2,000 was never the story. The story is that budget stopped being a usable excuse.

Build the problem inventory before the capability arrives