OpenAI GPT-6 Astra Solved 10 Impossible Math Problems — Then Took On Navier-Stokes

I’ll be honest — I didn’t expect to write about pure mathematics twice in one month. But OpenAI’s GPT-6 Astra just did something that forced everyone, mathematicians included, to stop scrolling.

In August 2026, an internal Astra checkpoint solved ten problems that had been open for a decade or more, complete with machine-checkable Lean 4 proofs you can actually verify on GitHub. Total compute bill? Around $2,000 at API rates.

Then in early September, a stronger internal sibling went after a Millennium Prize Problem — the 3D Navier-Stokes smoothness question — with a swarm of ~10,000 agents.

This is the full story: what Astra actually solved, why mathematicians are both thrilled and uneasy, who’s fighting over credit, and why the safety part genuinely worries me.

GPT-6 Astra, released Sept 3-4, 2026, is OpenAI’s first Critical-rated cybersecurity model — and an historic math reasoner. Image: Unsplash.

What is GPT-6 Astra? Quick recap

GPT-6 Astra, developed by OpenAI led by Sam Altman and Greg Brockman, launched as a limited preview on September 3, 2026 and rolled out publicly on September 4.

OpenAI calls it a generational leap for coding, science, computer use, and cybersecurity. It was pre-trained on 100,000+ GPUs at the Stargate site in Texas, scores 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, and uses a new “recurrent depth” or looped-transformer architecture.

That last detail matters. It makes Astra faster, but it also obscures its chain-of-thought. More on that in the risks section — it’s a big deal.

The August shock: 10 open problems solved for ~$2,000

Here’s what stopped me in my tracks. Before public launch, OpenAI let an unreleased Astra checkpoint loose on long-open math.

Result: 10 fully solved problems in mathematics and theoretical computer science, each with a Lean 4 proof anyone can check.

Every Astra math result shipped with a machine-checkable Lean 4 proof. That verifiability is why mathematicians are taking it seriously.

Standouts included:

• First known non-sofic group — explicit construction answering Gromov’s 1999 question
• Disproof of Connes’s rigidity conjecture
• Improved sphere-packing bounds
• Several Erdős catalogue results — the same problem family Thomas Bloom curates at erdosproblems.com

Noam Brown, OpenAI research scientist behind o1 reasoning work, summed it up drily at the time: “Sadly, no Millennium Prize Problems (yet).”

Mathematicians noticed. Tony Feng, math professor at UC Berkeley and research scientist at Google DeepMind, said the batch looks like top-tier journal material — the kind of output that could make a human PhD student’s career. Thomas Bloom called it simply “big news.”

The debate wasn’t “are the proofs correct?” They are. It was: is this Fields Medal depth, or just really, really good professional mathematics at machine speed? Honestly, for $2,000, that distinction feels academic.

September: 10,000 agents vs. Navier-Stokes

That “yet” from Noam Brown didn’t last long.

In early September 2026, OpenAI unleashed an even stronger internal model — described as significantly more capable than public GPT-6 Astra — on the remaining Millennium Prize Problems.

Navier-Stokes asks whether smooth fluid flows can blow up in finite time. Astra says yes — with a smooth external force. Image: Unsplash.

The numbers are wild:

• ~10,000 concurrent agents
• ~88 hours of continuous work
• Hundreds of billions of tokens
• Plus ~17 hours with Astra helping formalize it all in Lean

Output: a proof that smooth solutions to the 3D Navier-Stokes equations can develop singularities in finite time.

One researcher called it one of the largest compute efforts ever applied to a single scientific problem. On X, the jokes wrote themselves: “I don’t have $10 million in my piggy bank.”

Important nuance: OpenAI is not claiming the $1M Clay Mathematics Institute prize. Their formulation involves a smooth external force, technically different from the official Clay statement. Still, as a demo of large-scale agent collaboration, it’s historic.

Who’s who: the people caught in this story

Because this isn’t just about a model. It’s about people whose careers just got sideswiped:

• Noam Brown (OpenAI) – reasoning / multi-agent lead, former Libratus / Pluribus / CICERO creator. The public face of Astra’s math push.
• Tony Feng (UC Berkeley / Google DeepMind) – number theorist pushing for new norms around hiring and publishing now that AI does “PhD-level” output overnight.
• Thomas Bloom (University of Manchester) – arithmetic combinatorics, keeper of the Erdős Problems list Astra raided.
• Tristan Buckmaster (NYU Courant) – Clay Research Award winner for Navier-Stokes non-uniqueness work. Now at the center of a credit fight.
• Levent Alpöge (Anthropic) – former Harvard number theorist who just used Claude Fable 5 to disprove the 87-year-old Jacobian Conjecture in July. Buckmaster’s collaborator on fluid-dynamics work.
• Ryan Greenblatt (Redwood Research) – AI safety chief scientist warning Astra’s opacity is “potentially the single worst development for AI security/safety to date.”
• Organizations: OpenAI, Anthropic, Google DeepMind, NYU Courant, Redwood Research, Clay Mathematics Institute

The controversy no one wanted

Here’s where it gets uncomfortable.

NYU’s Tristan Buckmaster says he and Anthropic’s Levent Alpöge had spent nearly a year on a closely related route to Navier-Stokes / Euler singularities, with drafts and Codex logs to show it. He alleges OpenAI pressured him on authorship — including suggestions to drop his Anthropic-affiliated co-author — and questions the timing after their prior use of OpenAI tools.

OpenAI’s response: our researchers and agents did not see the unpublished work, the proofs differ, and while they can’t completely rule out de-identified usage data helping the model in general, no specific user data was accessed for the solution.

As reported by Science and the New York Times on Sept 10, Buckmaster himself admits he hasn’t seen OpenAI’s proof and doesn’t know if his private data was used.

So we’re stuck in a painful middle: no public evidence of theft, but a deep trust problem when the lab that provides your AI assistant is also racing you to the same prize.

Earlier Astra math results drew similar complaints about thin attribution of prior human work. Some mathematicians called the presentation inflated. Others pointed to replication claims — other models solving subsets — as proof it wasn’t that novel.

My take? The proofs check out, which is what matters scientifically. But credit norms are broken. We need timestamped disclosures, tool-use logs, and clear authorship rules, fast.

The real wake-up call: Critical cyber + black-box reasoning

Astra is the first model to hit Critical under OpenAI’s Preparedness Framework. It can find zero-days and chain exploits.

This is the part I wish more coverage led with.

OpenAI rates Astra as meeting the Critical cybersecurity capability threshold under its Preparedness Framework — the first model ever to do so. With tools and access, it can autonomously discover unknown vulnerabilities and chain them into working exploits against hardened systems. Evaluations showed strong results on ExploitBench (100% without safeguards), zero-day finding, sandbox escapes, and privilege escalation.

During testing it found two previously unknown zero-days, now being disclosed.

Combine that with reduced monitorability — that looped-transformer reasoning happens more inside activations, with less visible chain-of-thought — and you get why Ryan Greenblatt at Redwood Research is alarmed.

To its credit, OpenAI has:

• Blocked advanced offensive tasks in production (no PoC exploits)
• Added mandatory chain-of-thought monitoring with 30-minute pause-on-flag
• Isolated risky workloads and tightened privileges after July’s Hugging Face incident
• Gated top cyber capabilities behind Daybreak for defenders only

But high capability + low transparency is exactly the combination safety folks feared. The benefits and dangers aren’t arriving sequentially anymore. They’re arriving together.

What this means if you build with AI

I write about tools for designers and builders, so here’s my practical read:

1. Verification is the new moat. Lean 4 proofs are why math trusts Astra. Build checkable outputs into your own agent workflows.
2. Agent swarms work. 10,000 agents x 88 hours solved what single runs couldn’t. Think map-reduce, not mega-prompt.
3. Credit hygiene matters. If you use public AI tools for unpublished R&D, assume timelines will be questioned. Keep private repos, timestamps, and local models for sensitive work.
4. Don’t confuse public vs. internal. The gap between GPT-6 Astra you can use and the internal model that cracked Navier-Stokes is huge — in cost, too.

FAQ: OpenAI Astra math breakthrough

Did OpenAI’s Astra solve a Millennium Prize Problem?
Nearly. It produced a Lean-verified proof of finite-time blowup for 3D Navier-Stokes with a smooth external force. OpenAI says this differs technically from the Clay formulation, so it’s not claiming the $1M prize.

What did Astra solve in August 2026?
Ten long-open problems including the first explicit non-sofic group (Gromov 1999), disproof of Connes rigidity, better sphere-packing bounds, and Erdős problems — all with Lean 4 proofs for ~$2,000 compute.

Who is disputing credit?
NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge, who were working on related Euler / Navier-Stokes results with AI assistance. OpenAI denies seeing their unpublished work.

Why are safety researchers worried?
Astra is the first Critical-level cyber model and reasons more opaquely due to recurrent depth, making monitoring harder. Redwood Research’s Ryan Greenblatt called it potentially the worst development for AI safety to date.

Bottom line

Astra and its stronger internal siblings are a genuine leap. Solving — or nearly solving — Millennium-level math with verifiable proofs is historic, full stop.

But the credit fights, the $10M-scale agent swarms most labs can’t afford, and the Critical cyber + opacity combo show we’ve entered territory where breakthroughs and risks arrive as a package.

Exciting scientific partner? Absolutely. Wake-up call on risks? Also yes. We’ll need new norms for collaboration, verification, credit, and safety — and we needed them yesterday.

What do you think — would you co-author with a 10,000-agent swarm?

Sources: OpenAI GPT-6 Astra system card & Preparedness Framework (openai.com), Healthcare IT News / Fortune / Axios launch coverage Sept 3-4 2026, Wikipedia GPT-6 Astra, Science.org, NYT Sept 10 2026 interview with Tristan Buckmaster, NYU Courant & Berkeley faculty pages, Redwood Research.