OpenAI’s Astra just proved 10 long-standing math and science theorems. The tokens cost $2,000.

OpenAI shared a big research update this week, announcing that an internal version of Astra, its next frontier model, created machine-verified proofs for 10 long-standing problems in mathematics and theoretical computer science — all for roughly $2,000 in GPT-5.6 Sol API tokens.

That pricing is probably the most significant aspect of Monday’s announcement because it shows that the company is asking users to consider what a new discovery might cost once that model already exists.

For the first time, OpenAI gave us a rough idea of what that amount of AI reasoning might cost.

It’s important to note that the work isn’t finished, since the results require further review by human mathematicians. For developers, though, the biggest takeaway is that OpenAI says the research process produced 10 new results, and, for the first time, gave us a rough idea of what that amount of AI reasoning might cost.

OpenAI has not released the internal model or announced what it will cost to use. The $2,000 estimate is based on what the same number of tokens would cost at GPT-5.6 Sol’s API rates.

An internal version of our next major model produced 10 new results on long-standing open problems in mathematics and theoretical computer science, using roughly $2,000 worth of tokens at GPT-5.6 Sol API rates. pic.twitter.com/4cgowmPOpY

— OpenAI (@OpenAI) August 3, 2026

That rough comparison doesn’t account for the cost of training the model, running the surrounding research process, or having experts choose the problems and assess the answers. Yet, even with those limitations, the comparison is useful for developers.

Research labs cannot easily plan around claims that a model is ‘good at mathematics.’ But they can plan around an inference budget.

Inference over training costs

Research labs cannot easily plan around claims that a model is “good at mathematics.” But they can plan around an inference budget. If a model capable of frontier reasoning eventually becomes available through an API, teams could decide how much they are prepared to spend investigating a conjecture, testing a possible proof, or searching for a better bound.

The costs that make the biggest difference to the companies developing the systems typically come from power requirements to the billions of dollars needed to build the next generation of models.

And yet, they aren’t as important to a research lab that may eventually access the finished model through an API. For that lab, the relevant question revolves around how much one serious attempt at a problem will cost to run.

That’s the tone captured by OpenAI’s estimate. A startup would not need its own frontier model or a warehouse filled with GPUs because it could buy access to a trained model’s reasoning, just as companies now rent computing power instead of building their own data centers.

No guarantees, just opportunity

Of course, spending more money doesn’t guarantee a breakthrough. A model could burn through thousands of dollars in tokens and still come up empty.

But that uncertainty is already part of research and using an API wouldn’t change that. It would, though, give researchers another valid opportunity to pursue an answer.

Reasoning budgets for research

OpenAI researcher Noam Brown says the model used a relatively modest amount of test-time compute and that larger reasoning budgets remain available.

And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet).

But also, we didn’t spend a lot on each problem. It’s possible to push test-time compute much further.

— Noam Brown (@polynoamial) August 1, 2026

The model did not solve any of the Millennium Prize Problems, but OpenAI appears to believe it has not reached the limit of what more inference could produce.

Still, developers and research teams may soon need to decide how much a difficult problem is worth. A university might spend several thousand dollars exploring a mathematical question, while a pharmaceutical company could justify spending far more if a model helps narrow the search for a new drug.

What OpenAI has shown is that frontier reasoning can be discussed like any other computational workload.

Still, developers and research teams may soon need to decide how much a difficult problem is worth.

Democratizing frontier AI access

Developers already estimate what it costs to have a model generate code, analyze a database, or process a collection of documents, and research labs may eventually make similar calculations, which would change who can experiment with frontier systems. Building them may remain the domain of a few wealthy companies, but using them could become an API expense available to smaller research teams.

The post OpenAI’s Astra just proved 10 long-standing math and science theorems. The tokens cost $2,000. appeared first on The New Stack.