Before GPT-6: The 10 Math Problems OpenAI's Astra Solved, the Lean Proofs, and the GPT-5.7 Rumor

OpenAI has a model nobody can use. It shares its name with a Google project. And Sam Altman showed it to politicians in Washington before he showed it to you.

The name is Astra. In the last week of July, OpenAI published a report saying an internal version of Astra solved 10 open problems in mathematics, quantum complexity, and theoretical computer science. That same day, The Information reported that Altman gave a briefing on the model in DC. The internet connected the dots and read "GPT-6 incoming."

Slow down. Here we separate what OpenAI actually put on paper from what's thread gossip. This post is a tracker: every claim with a date and a source, updated as things heat up. Because with this kind of news, half of what's circulating is smoke.

TL;DR

• What it is: Astra is OpenAI's "next major model family," according to the company itself. Still unreleased, in testing.
• The fact that blew up: an internal version solved 10 open math problems, with proofs formalized in Lean (machine-checkable).
• What's still rumor: the commercial name (GPT-5.7? GPT-6?), the release date, and a good chunk of the specs.
• Official source: Ten advances in mathematics and theoretical computer science (OpenAI).

Update — September 3, 2026: Astra came out of stealth and launched as GPT-6 Astra. If you landed here wanting to know what the model is, what it costs, and how it stacks up against Claude Fable 5.1 and GPT-5.6 Sol, the definitive post is the GPT-6 Astra vs Fable 5.1 vs GPT-5.6 Sol comparison. This piece stays up as a record of the story of the 10 math problems that preceded the launch.

Update — August 4

The tracker is still live. Three things have changed since publication:

• The list of 10 got more concrete. Coverage from August 2 highlights two results that hadn't shown up by name in the first stories: the refutation of Connes' rigidity conjecture (operator algebras) and solutions to several problems posed by Paul Erdős — problems whose central results hadn't seen progress "in at least a decade, in most cases much longer" (BleepingComputer).
• GPT-5.7 entered the naming race. The same coverage now puts three candidates on the table: GPT-5.7, GPT-6, or a brand-new name. Still undecided — the table further down still holds.
• The technical community reacted — with a raised eyebrow. The report hit the front page of Hacker News (~390 points), and the skepticism there complements the mathematicians': the ~$2,000 compute cost (a number Noam Brown confirmed in the thread itself) suggests very well-orchestrated brute-force search, not necessarily a leap in model capability. And some compared the proofs to the four color theorem: a proof by construction that the machine verifies, but that doesn't leave a human understanding why.

What Astra is — the confirmed part

Let's go through what has an official stamp or a named source.

OpenAI named the model publicly. In the report on the math advances, the company refers to Astra as its "next major model family". This isn't an anonymous leak — it's OpenAI writing the name down.

Noam Brown, a test-time reasoning researcher at OpenAI, confirmed on X: an internal version of Astra "solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science" and would be "an important step toward scientific reasoning." He also admitted, honestly, that attempts on other problems failed — which already helps calibrate the hype.

The Information reported (behind a paywall, so it's a secondary source citing the primary one) that Astra is a multi-agent system trained for long-horizon tasks — problems the model chews on "for hours or days," instead of the short tasks current models handle. And that Altman previewed the model for officials in Washington.

If you follow the blog, this plot is familiar: it's the same pattern as the GPT-6 tracker — Altman in DC, a government briefing, and the community trying to guess whether the next model already has a name. The difference is that this time the name showed up.

The 10 math problems — what was verified

Here's what can be stated based on OpenAI's report. The results cover very specific fields:

• Sphere packing in high dimensions — new upper bounds on density, reaching the Cohn–Elkies threshold.
• Coding theory — exponentially improved bounds on the maximum size of binary and spherical codes at a given minimum distance.
• Group theory — a construction establishing the existence of non-sofic groups, a central open question in the field.
• Operator algebras — the refutation of Connes' rigidity conjecture (detailed in the August 2 coverage).
• Erdős problems — solutions to several problems posed by Paul Erdős in combinatorics (same).
• Plus results in lattice-based cryptography, quantum complexity, and extremal combinatorics.

Two points that keep this post from turning into a press release:

First, there's real verification. The proofs come with Lean certificates — a formalization a machine can check. That's a lot stronger than "the model said it proved it." A mathematician doesn't have to take the LLM's word for it; they run the certificate.

Second, it was expensive. The total tokens to reach those 10 solutions would cost about $2,000 at GPT-5.6 Sol API prices — a number that started as a community estimate and that Noam Brown himself confirmed in the Hacker News thread. This isn't a model that "thinks" for free. It's expensive compute solving a hard problem.

The mathematicians' healthy skepticism

When a lab says "our AI solved a math problem," the right reaction is to raise an eyebrow. And the mathematicians themselves did.

Thomas Bloom, a mathematician at the University of Manchester, called the results "big news" — but rejected the narrative that AI is replacing mathematicians. His argument is precise: the model uses "more than a century of mathematical theory" and was built by mathematicians. The machine didn't invent the field; it navigated a field humans built and found a new construction inside it.

That's the difference between "AI discovered new math" (headline) and "AI found a valid construction in a giant search space that humans defined" (reality). Both are impressive. It's just that the second one is true and the first one is marketing. We've seen this movie before with GPT-5.6 and the supposed 'new math'.

Is Astra GPT-6? Fact vs. rumor

This is the question driving the clicks, so let's be explicit.

The honest read: Astra exists, has a real and verifiable result, and is OpenAI's next bet. Whether the commercial name will be GPT-6, GPT-5.7, or just "Astra," nobody on the inside has confirmed. Anyone writing "GPT-6 is Astra" today is guessing.

What to do if you build on the OpenAI API

News about an unreleased model doesn't change your code today. But you can position yourself:

• Don't rewrite anything right now. Astra has no public API, no pricing, no SLA. Migrating your architecture over a headline is like the folks who switched models the day the price cut was announced and then had to redo the math.
• Keep an eye on the "long tasks" vector. If Astra really delivers an hours/days horizon with multi-agent, the real impact isn't in chat — it's in agentic pipelines that today you break into short steps because of context limits and reliability.
• Formal certificates are the trend that matters. Lean proofs point to where this is going: LLM output you verify, not output you trust. It's worth bringing that mindset into your own domain (tests, contracts, schemas) even without Astra.

Quick FAQ

Is Astra available yet? No. It's in internal testing. No public API, no pricing, no confirmed release date.

Does it have anything to do with Google's Project Astra? No. Just a naming coincidence. Project Astra is Google DeepMind's multimodal assistant. OpenAI's Astra is a family of reasoning models. Same name, different companies and projects.

Are the 10 math proofs trustworthy? The proofs come with Lean certificates, which are machine-checkable — that's strong. But "solved 10 problems" isn't "revolutionized mathematics"; mathematicians stress that the model operates on theory built by humans, and part of the technical community reads the results as expensive, well-orchestrated search, not as a leap in capability.

Does this confirm GPT-6? No. OpenAI hasn't confirmed the commercial name. Astra could become GPT-6, GPT-5.7, a variant, or keep the codename. Anything beyond that is speculation.

Conclusion

Astra is the kind of news that separates people who read sources from people who read headlines. The fact — 10 open problems solved with machine-checkable proof — is genuinely big and deserves attention. The rest — commercial name, date, "it's GPT-6" — is fog that will clear over the next few weeks.

This post is a tracker: I'll keep updating it as OpenAI confirms (or denies) the pieces still up in the air. In the meantime, if you want the other thread of this story — the launch everyone's waiting for — follow the GPT-6 tracker, where we do exactly this work of separating the confirmed from the guesswork.