On August 1, 2026, OpenAI for the first time officially confirmed the name Astra, its “next major model family.” To mark the announcement, the company released something unusual: a research report claiming an internal version of Astra had solved ten open problems in mathematics and theoretical computer science — problems where mathematicians had made no progress for at least a decade, and in most cases much longer.
The tokens that generated all ten solutions would have cost about $2,000 at Sol API rates. Each proof was formalized in Lean, producing machine-checkable certificates of correctness. The company published the full proofs, the reasoning walkthroughs, and a 249-page manuscript on GitHub.
Is this real? Yes, as far as the material goes. Is it as big as it sounds? That is the interesting part — and the part most headlines get wrong.
What OpenAI Actually Announced
Astra is a new model class alongside OpenAI’s existing Sol, Terra, and Luna families, according to The Information’s reporting, which cited three people familiar with the plans. CEO Sam Altman has already demonstrated the model to politicians and regulators in Washington, D.C. It has not shipped, and there is no release date. It has not even been decided whether it will be marketed as GPT-6 or as a variant within the GPT-5 line — something like GPT 5.7.
What OpenAI did publish is a math report. The claim: an internal version of Astra solved ten open problems in math and theoretical computer science. The areas involved run from high-dimensional geometry and coding theory to group theory, quantum complexity, lattice cryptography, and extremal combinatorics.
Here is what is genuinely remarkable. One proof establishes the existence of non-sofic groups, a major open question in group theory that had resisted progress for years. Others overturn the Connes rigidity conjecture, improve the upper bound for sphere packing in high dimensions — the first improvement since 1978 — and establish new circuit lower bounds for computing the permanent.
These are not toy problems. They are the kind of results a mathematician could spend a career chasing.
The $2,000 Detail That Makes People Uncomfortable
OpenAI says the tokens used to generate all ten solutions would have cost about $2,000 at Sol’s API rates. That is roughly $200 per solved problem.
Noam Brown, one of the researchers behind the test-time reasoning technology Astra uses, put it directly on X: the team also tried and failed to crack other major problems. “Sadly, no Millennium Prize Problems (yet),” he wrote. The Clay Mathematics Institute offers $1 million for each of the seven Millennium Prize Problems, and only one has been solved since the prizes were announced in 2000.
“But also, we didn’t spend a lot on each problem,” Brown added. “It’s possible to push test-time compute much further.”
That last sentence is the one to sit with. The point is not that $2,000 bought ten theorems. The point is that the cost structure of doing math is being rewritten, and we are at the very cheap end of it.
Why Mathematicians Care — And Why They Are Cautious
Thomas Bloom, a mathematician at the University of Manchester who runs erdosproblems.com, called the results “big news” on X. He considers them more significant than the counterexample to the unit distance conjecture published back in May. “Maybe not bigger than a proof of unit distance would have been, but in terms of constructions, this is big,” Bloom wrote.
He also pushed back on the framing that AI is replacing mathematicians. His argument: an AI that draws on more than a century of mathematical theory, was built by mathematicians, and was trained on everything mathematicians have ever written — calling that “replacing” mathematicians makes little sense.
That is a fair point, and it is worth keeping both halves. The work is real. The tools are real. And they are built on top of what humans did first.
There is also an honest caveat embedded in OpenAI’s own announcement. The company said its researchers helped prepare the papers and formalize the proofs, and that OpenAI takes responsibility for their accuracy. But the mathematical arguments themselves came from Astra. OpenAI argued that claiming human authorship for a proof generated entirely by AI would misrepresent both the system’s contribution and the nature of genuine human intellectual work — pointing to the Leiden Declaration on AI and Mathematics as a reference for how credit should be assigned.
Whether the broader mathematics community accepts all ten proofs is still an open question. Formal verification in Lean is a strong check, but the results still need to be reviewed, understood, and extended by human mathematicians.
What Makes Lean Verification Different From “AI Generated”
This is the part that separates Astra’s announcement from a thousand “AI solved X” headlines.
When a model produces an answer in a chat, you have to trust it. When Astra produces a proof, it also formalizes that proof in Lean — a proof assistant where every step is checked mechanically, line by line, by software that does not care how clever anyone thinks it is.
That is why the “machine-checkable certificate” part matters. It converts “the model claims this is a proof” into “a computer verified every step of this proof.” That is a different epistemic bar, and it is the reason mathematicians are taking this announcement seriously rather than dismissing it out of hand.
The distinction was also visible in how OpenAI framed the collaboration: humans worked with the model to turn its arguments into research papers. The model proved. The humans packaged, formalized, and took responsibility.
If you are still getting your bearings on where the major AI models stand right now, the ChatGPT vs Claude comparison gives you a solid baseline, and this week’s DeepSeek-V4-Flash launch shows the open-source side of the race.
What Astra Is Actually For
The math results are the demonstration. The product is something else.
OpenAI is building Astra to work on problems continuously for hours or days — coordinating multiple agents over long horizons instead of the short, single-shot tasks most current models handle. The stated use cases are complex projects and advanced mathematics. Chief Scientist Jakub Pachocki said on OpenAI’s official podcast last summer that the company wants systems that can plan, reason, and experiment over long time spans. By March 2028, OpenAI wants a fully autonomous AI researcher. As early as this September, it plans a system with research-intern-level skills.
Astra could be that system — or the model class that gets it started.
But this is also where the honest skepticism belongs. The original reporting from The Information flagged a known weakness in today’s agentic systems: they accumulate errors as workflows get longer, and they can fail to correct themselves as context grows. Multi-agent setups can even perform worse on tightly linked tasks like planning, because coordination overhead and compounding errors wipe out the gains. Long-horizon reasoning is precisely the part that is hardest to make reliable.
So: impressive math, untested long-run reliability. Both are true right now.
What This Means for You
If you are a researcher, this is the moment to stop assuming “AI can’t do math” and start checking what the proof assistants can already verify.
If you are a developer, Astra is worth watching for one reason: the same technology that produced these proofs is aimed at long-running, multi-step tasks — which is what most real-world agent work looks like. Whether it can avoid compounding errors over hours is the question that will determine whether agentic coding, research, and analysis actually scale.
If you are just following AI news, the healthy takeaway is calibration. Ten real theorems for $2,000 is a genuine milestone in scientific reasoning. It is not, yet, a proof of the Millennium Prize Problems. Noam Brown said it plainly: “Sadly, no Millennium Prize Problems (yet).”
Bottom Line
OpenAI confirmed Astra and backed the naming with real, machine-verified mathematics. The results are genuine, the cost is startling, and the long-horizon agent technology behind them is the actual story. It is a milestone — and an honest one, which acknowledges what it has not done yet.
Watch for three things: whether the mathematics community fully accepts the proofs, whether Astra can sustain long tasks without compounding errors, and what the U.S. government review — the first under the planned AI framework — decides before any public release.
None of those are decided yet. That is what makes this worth watching.