OpenAI Astra Solved 10 Math Problems

AI Summary

What this is: OpenAI says an internal version of its next model, Astra, solved ten open maths problems. This explains what happened and why the proofs matter more than the solutions.

Who it’s for: Anyone who saw the headline and wants to know whether it means anything for the AI tools they use.

The key idea: The results came with machine-checkable proofs, so nobody has to take OpenAI’s word for it. That is the part worth paying attention to.

Skip if: You want practical tool guidance. Try the best AI tools for beginners instead.

On 1 August 2026, OpenAI said an internal version of Astra, its next major model, had produced solutions to ten open problems in mathematics and theoretical computer science. The compute bill came to roughly $2,000.

Announcements like this normally sit unverified for months while specialists work through them. This one shipped with its proofs already machine-checked.

Table of Contents

What did Astra actually solve?

Ten problems spread across group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography and extremal combinatorics. Every one had been open for at least a decade. Most had been open far longer.

The headline result concerns something called a sofic group. A group, in this context, is a mathematical object describing symmetry and structure. A group is sofic if you can approximate its behaviour as closely as you like using finite systems, which is a way of saying you can get a handle on it with finite tools.

Mikhail Gromov asked the natural next question in 1999: is every countable group sofic? Essentially every group mathematicians work with day to day turned out to be sofic, but nobody could show it held universally, and nobody could produce a counterexample either. The question sat unresolved for 27 years.

Astra built one. The answer is no, not every countable group is sofic, and there is now an explicit construction demonstrating it.

Why the proofs matter more than the answers

AI systems assert things confidently that turn out to be wrong. That is the central problem with using them for anything where correctness matters, and it is why a lab announcing a mathematical breakthrough would normally be met with a long wait while humans checked the work.

OpenAI released the proofs written in Lean, a proof assistant. Lean checks each logical step mechanically, and it does not accept hand-waving. Work written in Lean either compiles or it does not.

The files went on GitHub alongside a 249-page manuscript. Lean marks any gap a mathematician has left for later with the keyword sorry. Across all ten formalised proofs, that count is zero, meaning the authors left no acknowledged gaps.

That shifts what you have to take on faith. You no longer need to trust OpenAI or wait on a reviewer, though you do still need to trust Lean itself and confirm the formal definitions say what they appear to say. Anyone can download the files and run them.

How have mathematicians responded?

Positively, with the caveat that the work is still being absorbed. Press reports quote Fields Medallist Timothy Gowers saying he would have recommended the proof for publication in a top mathematics journal without hesitation. That is second-hand reporting rather than a statement issued directly, so read it as a strong early signal rather than a formal endorsement.

Note what the comment is and is not. It judges the quality of one specific proof. It does not say mathematics is finished or that the field has been automated. A 249-page manuscript takes time to digest, and the community is still doing that.

Does this change anything for you?

Not today, and not directly. Astra has not shipped, so you cannot use it, and none of this changes what ChatGPT can do for you right now.

Two things here do matter for ordinary users, though, and neither is about maths.

The cost surprised people. Roughly $2,000 of compute produced results that had resisted specialists for decades. That figure covers the querying, not the training run behind it, so the accurate version is narrower than the headlines: once a model like this exists, asking it hard questions is cheap.

What carries over is the verification. This announcement holds up where others have not because the claims were made checkable. Most AI output has no equivalent. When a chatbot tells you something, there is no compiler to run it through, which is why you still have to check things yourself.

What this is not

Not AGI. Solving formalised maths problems is a narrow, well-defined task with a clear success condition. That is close to ideal conditions for a machine, and nothing like most human work.
Mathematicians have not been replaced. Humans posed these questions, built Lean, decided which problems mattered, and are now doing the interpretive work of understanding what the results mean.
Not a released product. This was an internal model. Public availability, pricing and capabilities are all unannounced.
Not fully settled. The Lean proofs verify the logic. Whether the formal statements capture the informal questions mathematicians care about is a human judgement, and that review is ongoing.

What AI can’t do here

Humans, not the model, decided which problems were worth solving. Gromov’s question mattered because mathematicians spent 27 years establishing that it did.
It cannot tell you whether a formal statement means what you think it means. Translating a real question into something a proof assistant accepts is where the human judgement sits.
It cannot supply the trust infrastructure. Lean exists because people built it over years, specifically so that claims like this one could be checked rather than believed.

The Big AI Stories, Explained Plainly

Human curated and edited, for beginners who want a finger on the pulse without the hype.

Free forever. Unsubscribe anytime.

Common questions about Astra

Can I use Astra?
Not yet. OpenAI described an internal version of the model. There is no release date, pricing or public access announced.

How do we know the proofs are correct?
They are written in Lean, a proof assistant that mechanically checks every logical step. The files are public on GitHub, so anyone with Lean installed can verify them independently.

What does a “sorry” count of zero mean?
sorry is the keyword Lean uses to mark a step someone has skipped. Zero of them means the authors left no acknowledged gaps. It does not remove the need to trust Lean itself, or to check that the formal statements match the questions they are meant to capture.

Did it really only cost $2,000?
That is OpenAI’s figure for the compute used to find the solutions. It excludes the cost of training the model itself, which is vastly larger.

Is this a bigger deal than previous AI maths results?
Formal verification separates it from earlier attempts. Previous systems produced results that then needed lengthy human checking, and machine-checkable proofs remove that bottleneck.

Does this mean AI can do research now?
It means AI contributed to results in a domain where success is formally checkable. Most research does not have that property, so this does not generalise as broadly as the headlines imply.

Google Wants a Dead Airline’s Data

How to Turn Off AI Everywhere

ChatGPT for Teens: Parent’s Guide