What Is OpenAI Astra? A Model That Ships Math Proofs

OpenAI named its next model family on August 1, 2026. There was no launch event. There was no benchmark chart.

Instead the company published a report on ten open problems in math and theoretical computer science, and pushed machine-checkable proofs to GitHub. The model is called Astra. The delivery format is the interesting part.

Key Takeaways

• Astra is OpenAI's next major model family, built to run for hours or days with several agents working one problem together.
• OpenAI announced it with ten Lean 4 proofs on GitHub rather than a benchmark table, so anyone can check the work by compiling it.
• Experts rate five of the ten proofs as likely correct so far. Several remain under review, and OpenAI has set no release date.

What Astra is

Astra is OpenAI's next major model family. Reporting describes a system that plans, revises, delegates and keeps working in the background for hours or even days on research, coding and science tasks, with several agents splitting one problem (the-decoder).

Two more terms matter for reading the announcement:

• Lean 4 is a proof assistant. You write a proof as code, and the compiler either accepts it or rejects it. There is no partial credit.
• A machine-checkable certificate is a proof in that form. Anyone can download it and run the checker. Trust in the author stops being the issue.

The naming is still open. OpenAI has not said whether Astra ships as GPT-6 or as a GPT-5 variant, and it has given no release date.

The ten problems

The results came from an internal version of the model. The problems span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and extremal combinatorics.

Four give a sense of the range (OpenAI):

1. Sphere packing. Improved asymptotic upper bounds on packing density, reaching the Cohn–Elkies threshold.
2. Binary and spherical codes. Exponentially stronger upper bounds for binary codes at every minimum distance.
3. Non-sofic groups. A construction of a non-sofic group, settling whether every group admits finite permutation approximations.
4. Connes's rigidity conjecture. A counterexample to the idea that certain groups are pinned down by their group von Neumann algebras.

Some context helps on the first one. The Cohn-Elkies bound is a linear programming method that gives the best known upper limits on packing density in high dimensions. It is sharp in dimensions 8 and 24, where it pinned the problem exactly (sphere packing in dimension 24). It is not sharp everywhere, and the gap between the bound and reality is where the open work sits (Cohn and Zhao on spherical codes). Moving an asymptotic bound to that threshold is a real result, not a tidy-up.

OpenAI put the token cost at roughly $2,000 at Sol API rates. That figure covers generating the ten solutions. It is not the cost of the research program behind them. Other outlets have summarized the same report, such as BitsMinds and AI Weekly.

Why proofs instead of benchmarks

Benchmarks have a credibility problem. A lab reports a score, and the first question is whether the test leaked into training. The second is whether the eval was tuned. Neither question has a clean answer from outside the lab.

A Lean proof sidesteps both. The Lean certificates on GitHub compile or they do not. You do not need to trust OpenAI's evaluation process, its prompt, or its sampling budget. You need a compiler.

That is a real shift in how a lab can make a claim. Compare the two formats.

Benchmark score	Lean certificate
Verified by the lab that ran it	Verified by anyone with the toolchain
Contamination is hard to rule out	Contamination is irrelevant to correctness
A number with no artifact	An artifact you can execute
Says the model is good at a test	Says one specific claim is true

There is a trade. A proof tells you the model did this, once, on this problem. A benchmark tries to tell you what the model will do next time. The proof is stronger evidence about a narrower thing.

Consider a hiring analogy. A benchmark score is a candidate's grade average. A Lean certificate is a work sample you can run yourself.

What the proofs do not settle

Here the coverage has been thinner than it should be, and this is the part worth slowing down on.

OpenAI says humans prepared the arguments into manuscripts using the same model, and the model then formalized each argument in Lean. So the pipeline is a collaboration, not an unattended run.

Expert review is also incomplete. Based on feedback so far, at least five of the ten attempts — problems 4, 5, 6, 9 and 10 — have a high chance of being correct. Several others remain under review. That is a good result. It is not ten solved problems, and the difference matters if you are trying to size the capability.

Formalization has limits too. A Lean proof shows an argument follows from its stated assumptions. It does not confirm the statement is the one mathematicians care about. Getting that translation right is human work, and it is where a formal result can still mislead.

What it means for everyone else

Very little changes this month. Astra is not available, and no date has been set.

Three second-order effects are worth watching.

• Verifiable artifacts as a claim format. Expect more labs to ship something checkable next to a score. Formal proofs suit math. In software the equivalent is a reproducible benchmark harness or a failing test that a patch turns green.
• Long-horizon agents get a real test case. A model that works a problem for days is the same architecture people are pointing at business tasks. Results there have been rough, as we found writing about the AI agent that ran a business and lost $447.
• Policy attention. Astra may be among the first models to use the federal prerelease framework, which we covered in our piece asking whether the government has to approve AI models. That framework is voluntary, and it grants no veto over a launch.

For example, if you evaluate AI coding tools, the useful question shifts. Not "what did it score on SWE-bench," but "can you hand me the artifact that proves it." That question is answerable. A score is not.

Frequently Asked Questions

What is OpenAI Astra?

Astra is OpenAI's next major model family, announced August 1, 2026. It is designed for long-running work, where several agents coordinate on one problem for hours or days.

Did Astra solve ten unsolved math problems?

It produced solutions to ten open problems, with Lean certificates for each. Experts rate five as likely correct so far. Others remain under review, so "ten solved" overstates the current position.

When is OpenAI Astra released?

There is no announced date. OpenAI has also not confirmed whether it ships as GPT-6 or as a variant of GPT-5.

Why does a Lean proof matter more than a benchmark?

A Lean certificate can be checked by anyone who compiles it, so training contamination and evaluation tuning stop mattering. A benchmark score depends on trusting the lab that produced it.

Did the model work unassisted?

No. OpenAI says humans prepared the arguments into manuscripts with help from the same model, which then produced the formal certificates.

The short version

Astra is unreleased, unnamed in version terms, and partly unverified. Treat "ten solved problems" as five likely plus several pending.

The lasting part may be the format. A lab published artifacts that strangers can check, instead of a number that only it could produce. If that becomes the norm for big claims, evaluating models gets easier for everyone outside the labs.