FIELD NOTES · AN INTERACTIVE BRIEFING · 30 AUGUST 2026, UPDATED 6 SEPTEMBER 2026
GPT-6 Astra, and the ten doors it opened
OpenAI spent August showing one strange, specific thing about GPT-6 Astra and almost nothing else: ten mathematical doors that had stayed shut for decades, standing open. On 3 September 2026 it shipped the model itself, with a price list and a cybersecurity restriction attached. This page walks the trail, checks every claim against its primary source, and lets you open the doors yourself.
The short answer: GPT-6 Astra shipped on 3 September 2026, settling both the name and the date that August spent arguing about. OpenAI lists it at a 1,050,000 token context window and 128,000 maximum output, priced at 10 dollars per million input tokens against 50 per million output. It is the first model OpenAI says reaches the Critical cybersecurity threshold in its own Preparedness Framework, and the version you can buy refuses the advanced cyber work that rating describes. The ten Lean-checked mathematical results that started the story still stand, and are still not peer reviewed.
10 doors to open 2 minutes or 20, your pace progress saves on this device
01 · THE RECORD
Four dated documents, then the model itself
Strip away the speculation and the pre-release record of the Astra model was four stones on a trail. In the last week of July, Sam Altman demonstrated it privately to United States senators and officials in Washington. On 1 August, OpenAI published Ten advances in mathematics and theoretical computer science, crediting "an internal version of Astra, our next major model". On 7 August it published a cyber-capability disclosure that paused parts of Astra's own development. And on 18 August it detailed the safeguards that pause bought. Then, on 3 September 2026, it shipped the model as GPT-6 Astra, and the speculation was replaced by a specification.
Everything else you read about GPT-6 Astra before September descends from those four documents, a handful of staff posts, and reporting on the private demos. This page cites the documents, not the descendants.
four stones, one trail
The whole verifiable Astra timeline fits on one line. Sources: OpenAI's own posts of 1, 7 and 18 August 2026; Washington reporting by The Information, 31 July.
02 · THE TEN DOORS
Ten problems, ten doors. Open them.
The 1 August release is why anyone knows the name. An internal version of Astra produced results on ten problems spanning high-dimensional geometry, coding theory, group theory, operator algebras, circuit complexity, quantum complexity, lattice cryptography and extremal combinatorics, per OpenAI's announcement. Where headlines said ai solves math problems, the paper itself is more precise: constructions, counterexamples and bounds, several on questions that had not moved in decades.
Each star below is one problem. Tap or click a star to open its door; a keyboard reaches every one. Open all ten and the page will notice. Your trail is saved on this device only.
Fields and results per the announcement and the 253-page manuscript. The map is illustrative; the mathematics is not.
03 · ONE DOOR, PROPERLY
The number that had not moved since 1978
Take one door and walk through it slowly. How densely can identical spheres pack, as the number of dimensions grows without bound? The best general exponent was set by Kabatianskii and Levenshtein in 1978 and stayed put for forty-eight years. The Astra manuscript improves it, and then closes the road behind itself: it proves the Cohn-Elkies linear-programming method, the main tool in the field, can never push past the new number.
The packing exponent finally moved
Tap either endpoint for its story.
A larger exponent means a stronger upper bound on how densely spheres can ever pack in high dimensions. The manuscript calls this the first improvement to the general exponent since 1978. Source: the paper, Chapter 1.
04 · THE SEAL
What the wax seal actually certifies
Every result ships with lean proofs, published in the open ten-proofs repository under Apache 2.0. The repository's own manifest records zero unfinished goals, only three standard axioms, Lean 4.32.0, about one week of formalisation wall time, and a review status of exactly two words: "agent-reviewed".
That last clause is the honest one. A compiling certificate proves the code's theorem; whether the code faithfully captures the fifty-year-old conjecture, whether the result is novel, and whether it matters remain human questions. On the Lean community's public discussion thread, Kevin Buzzard summed his reaction as excitement and optimism, reviewers called the code readable but not up to Mathlib standard, and one OpenAI engineer acknowledged a non-idiomatic pattern had been introduced by tooling during conversion rather than by the model.
05 · THE ARGUMENT
Nobody disputes the proofs. Everything else is disputed.
Four weeks on, no named mathematician has challenged a single proof's validity. The live fight is about credit and framing. Steven Miller told Scientific American the sphere-packing chapter leant on his 2016 work without credit; Francesco Fournier-Facio said key non-sofic ingredients existed in 2016 and 2019 papers and the original no-progress-for-a-decade framing was wrong. OpenAI softened the wording and revised the manuscript: the original ran 249 pages, the current file is a 253-page update dated 6 August that links its own predecessor.
The sharpest twist: Fournier-Facio, the same mathematician who criticised the framing, then posted an arXiv paper building on the Astra criterion. And a fortnight earlier, Shuoxing Zhou had posted an independent, concurrent counterexample to Connes rigidity, found with help from GPT-5.6 Sol.
"These results, and the other eight on the list, are extraordinarily impressive"
"disturbing to see the recent trend of proofs being abandoned at an intermediate stage of their developmental process"
"yes, I would rank this as bigger than the unit distance counterexample"
06 · THE PAUSE
The model so capable its maker slowed down
On 7 August OpenAI disclosed something no frontier lab had said about its own unreleased model: internal evaluations of Astra showed advances in agentic coding and cybersecurity strong enough that it could not rule out the Critical cyber threshold of its Preparedness Framework. That threshold's definition includes finding and building working zero-day exploits against hardened real-world systems without human help. The response, per the 18 August post, was to slow its own scaling.
That question closed on 3 September, and it closed in the direction OpenAI had flagged. The launch post states that Astra "meets the Critical threshold in cybersecurity under our Preparedness Framework", and the accompanying safety overview says it without hedging: "Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework." What ships is deliberately narrower than what was measured: OpenAI says the launch version will refuse advanced cybersecurity tasks such as writing proof-of-concept exploits, while secure code review and patching are allowed, with looser safeguards promised to vetted defenders through its Daybreak programme. A capability rating and a purchasable product are not the same object, and the distance between them here is the safeguard. Treat that as OpenAI's account of its own model, because that is what it is.
One conflation to clear while we are here: OpenAI's own agents did breach Hugging Face this month, and folkfox covered that incident when it happened. But Astra was not the culprit, per both the 7 August post and Noam Brown directly. The frightening detail runs the other way: the breach models were merely GPT-5.6 Sol scale.
07 · THE NAME, THE DATE, THE PRICE
It is GPT 6, and it opened on 3 September
For a month the honest answer was that nobody knew, OpenAI included. Two independently sourced reports, from The Information via The Decoder and from BleepingComputer, agreed the company had not chosen between GPT 6, a GPT-5.x label such as GPT 5.7, or a separate family name entirely. Anyone stating openai astra IS GPT 6 was, in August, ahead of OpenAI itself. It is not ahead any more. OpenAI announced the model as GPT-6 Astra on 3 September 2026 and began rolling it out the same day, first to a limited set of organisations, then to ChatGPT Plus, Pro, Business and Enterprise users and through the OpenAI API, Microsoft Azure and AWS Bedrock. The Spud precedent, rumoured all spring as GPT 6 and shipped in April as GPT-5.5, did not repeat.
The gpt astra release date question that dominated August searches resolved on the third. The specification arrived with it, and this is the part a buyer should read twice:
Two lines in that price list decide more than the headline rate does. Cached input costs a tenth of fresh input, so a system that reuses a stable prompt prefix pays 1 dollar per million where a system that rebuilds its context each time pays 10, and writing that cache costs 12.50. And a request carrying more than 272,000 input tokens is repriced across its whole length, at double the input and cache rates and half again on output. The million-token window is real; using all of it is simply not priced like using the first quarter of it. Where a workload sits either side of that line is an architecture decision with a number attached, which is the sort of thing our AI consultancy work exists to settle before the invoice does.
Context window, output ceiling and all four prices read on 6 September 2026 from OpenAI's own API model page for GPT-6 Astra, which also records a 30 April 2026 knowledge cutoff. Rollout and rating from the 3 September announcement. Prices are OpenAI's list rates and move without notice.
August had opinions about the date, priced hourly. They are worth keeping as a record of what the crowd believed:
The market said mid-September. It opened on the third.
Tap a bar for its story.
Probability that Astra would be publicly released by each date, per Polymarket's market rules, which required public availability and a credible identification with Astra. Snapshot taken 30 August 2026, 13:08 UTC, on $412,910 of traded volume, four days before the model actually arrived. The crowd's central expectation, the 68.5 per cent bar, sat a fortnight later than the real date: the market with the most money on the question still underpriced how soon Astra would land. Kept here as a record of what was believed, not as a live price: the live market moves hourly.
08 · THE 30-DAY WINDOW
First model through a federal door
Before the public knew the name, senators saw the model. In the last days of July, Altman demonstrated Astra in Washington: closed-door meetings with Senators Warnock and Moreno, Senator Warner scheduled, Treasury and Commerce secretaries on the itinerary, per reporting by The Information. TIME later described a demonstration of sixteen agents dividing a research-level maths problem, coordinating, and assembling a proposed proof.
The door it walks through is Executive Order 14409, signed 2 June: classified benchmarking decides which systems count as covered frontier models, the NSA Director makes the designation, and developers may give government up to 30 days of pre-release access, per the order's text. The order explicitly disclaims licensing and preclearance: voluntary, not a veto. Astra is expected to be the first model through it.
For anyone selling into regulated categories, this rhymes with ground folkfox walks daily: the same week a company slowed itself for safety review, its product demo toured the capital. Restraint, done in public, is positioning. Our cybersecurity practice makes that argument for a living.
09 · THE MURMURATION
What people are saying, answered
A month of discourse, checked against the record. Each claim below circulates on X in some volume; each verdict traces to a primary source cited on this page. Tap a claim to see its answer.
OpenAI's 7 August post says Astra was not involved, and Noam Brown repeated it plainly on 26 August: the incident was driven by models of roughly GPT-5.6 Sol scale. The true detail is more sobering than the myth: the breach did not need a next-generation model.
Undecided for a month, then decided. OpenAI announced the model as GPT-6 Astra on 3 September 2026. The reports saying the name was still open were accurate when written, and the Spud precedent, rumoured GPT 6 and shipped as GPT-5.5, simply did not repeat this time. Anyone who called it in August was right without being able to show why.
It did not, and OpenAI's Noam Brown said so directly. The ten are long-open research problems, several of them counterexamples; none carries a Clay Institute prize.
OpenAI's wording: the tokens to find all ten solutions would cost roughly 2,000 dollars in total at Sol API rates, a counterfactual price at another model's rates. It excludes the failures, the manuscripts, the formalisation and the humans. Even careful readers slipped: one prominent blogger rendered it as per-problem.
Zero journal or conference acceptances exist. The repository's own manifest labels review status agent-reviewed, and the community catalogue of Erdos problems still lists all three as open. Machine-checked and peer-reviewed are different things, and the difference is the story.
No named mathematician disputes a proof's validity. The real disputes are about novelty and credit: which prior work went uncited, and whether a decade-of-no-progress framing was honest. OpenAI revised the paper in response.
His publish-without-hesitation line was about May's unit-distance result, not the August ten. What he actually said about the ten: extraordinarily impressive, with careful caveats about what AI is and is not yet good at.
Written on 30 August, when no date existed anywhere in OpenAI's record and the market priced release that month at two per cent. Astra arrived four days later, on 3 September 2026. The guessers were right by luck rather than by evidence: the leak accounts naming internal checkpoint strings were never acknowledged, and the market holding the most money on the question still had the date too late.
10 · WHAT IT MEANS
Now the door is open
The gpt 6 release date question resolved on 3 September, and it resolved into the shape the trail had already made visible: a model built to work for hours or days, coordinate agents, and produce work whose checking becomes the bottleneck. The marketing consequence is the one this site keeps returning to: machines increasingly read, cite and act on your pages before humans do. That is the terrain of our search and answer-engine work, the delegation patterns in our Hermes Agent piece and its raid-night companion, and the vendor-trust questions in this week's advisory coverage.
Read both scoreboards, then ask what you can buy
Every laboratory publishes a table in which its own model wins, so take each of these as the vendor's own claim rather than as a ranking. On Agents' Last Exam, a professional-work agent benchmark, OpenAI reports GPT-6 Astra at 59.3 per cent against 55.5 for Claude Opus 5 and 53.6 for its own GPT-5.6 Sol, and says Astra needed roughly 65 per cent fewer output tokens to get there. On Terminal-Bench 4.0 the two vendors look at first like they disagree and on inspection do not. OpenAI reports Astra at 57.9 per cent against 55.8 for Claude Fable 5.1. Anthropic reports 60.9 per cent, comfortably ahead of Astra, but that score belongs to Claude Mythos 5.1, and Anthropic's own page says Fable 5.1 and Mythos 5.1 are the same underlying model, the gap between them reflecting tasks where Fable's cyber safeguards intervened. Where the two tables overlap they agree exactly: both put Fable 5.1 at 55.8 per cent and GPT-5.6 Sol at 37.3.
Which leaves a buyer a better question than who won. The 60.9 belongs to a model Anthropic restricts to its trusted access programmes, so it is a benchmark leader you cannot purchase; Fable 5.1 is the one generally available. Astra, meanwhile, ships restricted in precisely the domain it scores highest in. No independent third-party run has been published to arbitrate any of it. The number worth deciding on is the one attached to a model you can actually deploy, under the safeguards it will actually carry, and that number is rarely the number in the announcement.
Astra figures from OpenAI's 3 September announcement; Mythos and Fable figures from Anthropic's Fable 5.1 and Mythos 5.1 page. Both read 6 September 2026. Vendor-published throughout, and dated because this field moves in weeks.
TRAIL QUIZ · EARNED
✦ Trail complete. The quiet reader discount on our attention: earned.
Building for the readers that read at machine speed?
folkfox writes pages that answer engines cite and humans finish: sourced to primaries, honest about uncertainty, and quicker than the cheat sheets. The proof is the page you just walked.
Method note: every dated claim on this page was verified against the linked primary source on 30 August 2026, including OpenAI's posts (read directly), the manuscript PDF, the ten-proofs repository, both arXiv papers, the executive order, the named blogs, and the market API. The two X posts quoted were opened and read before quoting. Market prices move; ours are a dated snapshot. Art is Nano Banana Pro watercolour, animated by the page, never by the facts.