In six days OpenAI did two things.
On 3 September it launched GPT-6 Astra, and on its own launch page sits a line easy to miss between the charts: the model meets the "Critical" threshold for cyber capability under the company's Preparedness Framework.
On 8 September it published a resolution of the Navier-Stokes problem - one of the seven Millennium Prize Problems, open for roughly ninety years.
And the two are not the same model. That, as we will see, may be the most important detail in either of them.
This piece is based on figures OpenAI published itself. They have not been verified by an outside party, and that is itself part of the point.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
Part One: What the Launch Said About Cyber
The Preparedness Framework is the internal classification OpenAI built to rank the risk of its own models. "Critical" is the highest tier in it. The company is publishing here, voluntarily, that it has crossed the bar it set for itself.
| The benchmark | Astra | Prior generation |
|---|---|---|
| ExploitBench - turning known vulnerabilities into working exploits | 100% | 78.5% |
| ExploitGym | 42.4% | 30.3% |
| SRE-Bench - reverse engineering binaries, single attempt | 88.0% | 55.9% |
| SRE-Bench, within four attempts | 99.2% | 68.7% |
And the most important detail is not in the table.
Because exposure to a corpus of historical vulnerabilities could have flattered the results, OpenAI built an additional benchmark using vulnerabilities from the previous three months only. And during that evaluation, per the company, the model found and used two zero-day vulnerabilities that were not known to anyone.
The company said it is disclosing both to the software maintainers.
And beyond the benchmarks, the company wrote, expert-led assessments found that the model - run without production safeguards - "could use previously unknown vulnerabilities to achieve arbitrary code execution in hardened browsers and create privilege-escalation exploits for hardened operating-systems".
What the Company Did Block, and What It Plans to Open
This is the part that tempers the picture, and it deserves to be given in full.
In the launched version, the model refuses advanced cyber tasks - for instance building a proof-of-concept exploit for a vulnerability. What is permitted is the defensive side: secure code review, and patching.
But OpenAI writes explicitly that this is temporary. Through a programme called OpenAI Daybreak it plans to "expand access and roll out less restrictive safeguards in the coming weeks", to enable vulnerability validation, malware analysis and detection engineering.
This is the point at which it becomes a market matter rather than a laboratory one.
A capability that exists but is blocked changes nothing on anyone's balance sheet. A capability that opens gradually, on a stated schedule, does.
And it works in both directions at once: the same capability that lets a defender find a weakness before an attacker also lets the attacker find it - and the difference between the two is who gets access and when. OpenAI itself puts it this way, that frontier cyber capabilities help defenders find weaknesses faster, but also "make those weaknesses easier to exploit".
And the Side That Moves Sooner: Cost
While the headlines went to cyber, there is a second line on the launch page that will affect more companies and faster.
| The benchmark | Astra | The rival | The cost gap |
|---|---|---|---|
| Agents' Last Exam - professional tasks | 59.3% | Claude Opus 5 - 55.5% | about 65% fewer output tokens |
| Terminal-Bench 4.0 | 57.9% | Claude Fable 5.1 - 55.8% | about 63% lower cost per task |
| BenchCAD | 95.9% | Claude Fable 5.1 - 84.3% | about 86% lower cost |
| Terminal-Bench Science | 64.6% | Claude Fable 5.1 - 52.6% | about 31% lower cost |
(All figures as OpenAI presents them, in configurations it selected. Anthropic had not published a response to these figures as of writing.)
The pattern is consistent: a higher score at materially lower cost. And that is the thing that enters the accounts of real companies - not as a "capability" but as an expense line.
The company also measured time, not only money: on a computer-use benchmark, 72.6% at roughly 40 minutes per task against 65.7% at roughly 75 minutes for the prior generation. A higher score in about half the time.
Part Two: And Five Days Later a Millennium Prize Problem Fell
On 8 September OpenAI published that an internal system of its own had produced a resolution of the Navier-Stokes problem.
The Navier-Stokes equations describe the motion of fluids, and are used in aircraft design, weather forecasting and the study of blood flow. The open question was whether smooth flow can break down - that is, whether the equations can develop a singularity, a state in which speeds grow without bound in finite time, even when motion starts smoothly and even with viscosity acting to smooth it.
And the answer obtained is yes. Per OpenAI, the system proved that a smooth fluid at rest can develop a singularity in finite time, with a smooth force applied and its energy remaining finite throughout.
This is a disproof rather than a proof - a negative answer to whether smoothness is preserved. In the official formulation it establishes statement "C" and also "D".
The physical picture: a vortex spiralling inward and progressively elongating, like spaghetti. The central region shrinks and accelerates - but exactly so that the energy stays finite.
This does not mean the equations are wrong - it means there are conditions under which their continuum approximation breaks, and from that point each particle has to be tracked individually.
And It Was Not Astra That Solved It
This is the detail most coverage missed, and it is material.
In the company's words: "To solve the Navier–Stokes problem, we used an internal model that is significantly more capable than GPT-6 Astra."
| What is known about the model that solved it | |
|---|---|
| When its training began | 28 August 2026 |
| Its status | per the company, training continues and performance keeps improving |
| Has it been released | No |
| Astra's role here | verification only - 17 hours of formalisation in Lean |
So the model billed at launch as the world's most intelligent stopped being the frontier within less than a week - not because of a competitor, but because of another model at the same company that has not been released.
And that changes how every number above should be read: they describe what is already sold, not what already exists.
How It Actually Worked
| The process | |
|---|---|
| When the effort began | 1 September, per the company following rumours that two Millennium problems had been resolved |
| The structure | groups of coordinating agents, with access to a cached version of the internet and the ability to run code |
| Size of the successful group | about 10,000 concurrent agents |
| The method | different groups were given different statements - some leading to a proof, some to a disproof |
| When the solution arrived | Saturday 5 September - about 88 hours after the first agents launched |
| Consumption across all problems | 4.9 million messages, about 300 billion output tokens |
| Navier-Stokes alone | 2.7 million messages, about 130 billion output tokens |
And another problem fell before Navier-Stokes. OpenAI also asked the system to try "easier" problems, and one of them - the regularity problem for the Euler equations in its unforced version
- was resolved, to their surprise, by about a hundred agents in roughly 50 hours. That success is what prompted them to shift resources to Navier-Stokes.
And per TechCrunch, 300 billion output tokens equate to roughly $22.5 million of compute if charged at current Astra rates. (That is TechCrunch's calculation, not an OpenAI figure. It assumes a particular rate and is theoretical - the company ran this on its own infrastructure and did not pay itself list price.)
But even as a rough estimate it says something: a proof considered out of reach for ninety years acquired a price tag, and that tag is the order of magnitude of a university department's annual research budget.
And the Dispute Attached to It
And here one has to stop, because the story is incomplete without it.
On the same day, Tristan Buckmaster, a mathematics professor at NYU, published three proofs - with Levent Alpöge, a mathematician employed by Anthropic (per the reporting, not on the company's behalf). Among them a preliminary finding on Navier-Stokes.
What Buckmaster claims, from the statement he published:
- That he and Alpöge discovered that "information about our progress had been passed to OpenAI".
- That when they contacted the company they were told a full proof had already been achieved - and that when they asked when the research began and how much human input was involved, the answers became evasive.
- "Eventually, it was agreed that [the first prompt] had been sent in the past few days, after information about our work had reached OpenAI."
- And on the direction: "The route to the Clay problem through a smooth force, options c and d in Fefferman's statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement."
And a graver claim: he says OpenAI mathematician Sébastien Bubeck asked him to remove Alpöge's credit as part of a proposed compromise, and that when he pushed to make the dispute public, Bubeck replied: "Why would you ruin your career?"
Bubeck published a response disputing parts of the account, but apologised explicitly for that line - saying it was an extremely poor choice of words, the opposite of what he was trying to convey, and that he retracted it on the spot.
And OpenAI does not deny the timeline - it confirms it. Its post states that the effort began on 1 September "after hearing a rumor". And on access to the work itself, in its words:
"We (the researchers and the agents) did not see any of their work through any means until they released it publicly - in particular, no specific user data was accessed in order to solve this problem. While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
The company stresses that the proofs differ significantly, and that Buckmaster and Alpöge's result concerns forced Euler while the agents' Euler result was for the unforced version. It recognises the priority of their work and congratulates them on it.
And a final detail: Buckmaster and Alpöge themselves worked primarily on OpenAI's Codex. The company reserves the right to train models on Codex interactions, and users are able to opt out.
And OpenAI said it does not intend to claim the Millennium Prize.
And What About the Cyber Stocks
That is the obvious question, so I checked it - and the answer is that you cannot see the event in the prices.
The rollout began on 3 September. That day CrowdStrike rose 5.68% and Palo Alto 1.06%. But neither can be attributed to the launch: CrowdStrike was still digesting a quarterly report from the week before, and Palo Alto was one day past a 9.28% fall on its own results.
In other words those two readings are too noisy to learn anything from - and that is the answer, not the absence of one.
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
What I take from this week is not the cyber and not the mathematics, which I am not able to judge - but three things that can be seen.
First: the model sold to the public is not the frontier. Astra's launch used the words "most intelligent in the world". Five days later the same company published that its most important result was produced by a different, more capable model whose training began on 28 August and is still running. Anyone gauging the pace of progress from the products on the market is looking at a picture that is late by definition - and that applies to every number in this piece too.
Second: what was bought here is compute, not insight. Ten thousand agents, 88 hours, 300 billion tokens. This is not genius appearing - it is a wall broken by force. And if problems considered unsolvable fall to a flood of compute, the advantage moves to whoever has more compute and not necessarily to whoever has the idea.
And that is exactly what Buckmaster says happened to him. I do not know how to settle the dispute and will not pretend to - both sides published accounts, Bubeck apologised for one sentence, and OpenAI confirmed the timeline while rejecting the conclusion. But the claim itself describes a structure that is possible, and both companies concede an indirect path cannot be entirely ruled out.
And the third, which runs through the whole piece: a commercial company preparing for a listing published voluntarily that its product crosses the highest risk threshold it defined, and launched it in the same breath - then published a larger result from a model nobody has seen. Nobody publishes "Critical" for the headline, so I do not dismiss it. But every number here belongs to the party selling the product, in configurations it chose, against rivals who did not respond.
That does not make them wrong - it does mean they remain a claim and not yet a finding.
(It is important to stress: this is my personal opinion only, and nothing herein constitutes a recommendation to take any action. Ilan Abramov, the writer, or parties on his behalf may hold positions in the securities mentioned, and those holdings change from time to time.)
