After GPT-6 Astra: what OpenAI has actually said about its next model
On 2026-09-08, in the post announcing its Navier-Stokes Millennium Prize result, OpenAI wrote that it used "an internal model that is significantly more capable than GPT-6 Astra", and that "since August 28 we have been training a new internal model... This model's training is ongoing and its performance continues to improve." So the next model's existence is officially confirmed, not rumoured. What the same post does not contain: a name, an API model id, a release date, or a price. This page separates the official wording, the checkable timeline, and the preparation that is actually worth doing today.
Updated 2026-09-10
Four things that are settled (each with official wording)
Training start date
Official wording: "Since August 28 we have been training a new internal model." This is the only exact date OpenAI has given for this model, and it anchors the whole timeline.
Training has not finished
Official wording: "This model's training is ongoing and its performance continues to improve." Training still running means there is no callable version today. Anything claiming it is already usable is not the official position.
Relative to GPT-6 Astra
Official wording: "we used an internal model that is significantly more capable than GPT-6 Astra." That relative phrasing is all OpenAI gave; the chart in the post carries no citable scores, so this page gives no "how much stronger" number.
Nothing published at all
The post contains no name, no API model id, no release date, no price and no context window. Any name you see elsewhere is not official, and this page does not repeat one.
What this model is, and what it is not
It is an internal OpenAI model, not a product. In the 2026-09-08 post, OpenAI used it (more precisely, a multi-agent system powered by it) to produce a resolution of the Navier-Stokes existence and smoothness problem, together with a formalization in Lean. The entire public description of the model amounts to three points: training began 2026-08-28, training is ongoing, and it is significantly more capable than GPT-6 Astra. It is not a point release of GPT-6 Astra, and OpenAI has not placed it in any shipped product line. It is also not "GPT-7" - a name OpenAI has never used. The most common mistake when reading about this topic is treating capability claims in a research post as specifications of an upcoming product. The post describes an internal evaluation, not a launch.
The checkable numbers in the 2026-09-08 post
The group that produced the Navier-Stokes resolution involved "on the order of 10,000 concurrent agents". The agents reached their resolution on Saturday 2026-09-05, about 88 hours after the first agents were launched; Lean formalization and verification then took an additional 17 hours, and that step ran on GPT-6 Astra. Across all attempted problems the agents sent 4.9 million messages and used about 300 billion output tokens; the Navier-Stokes problem alone accounted for 2.7 million messages and roughly 130 billion output tokens. OpenAI also states plainly that it does not intend to claim the Millennium Prize for the result. The value of these numbers is not the scale itself but the yardstick they give you: they describe a one-off internal research effort, not something anyone can reproduce through an API.
Timeline (all three entries from the same official post)
OpenAI begins training the new internal model, which it says shows "unprecedented" performance on internal benchmarks including mathematics.
The agents produce the Navier-Stokes resolution, about 88 hours after launch. Lean formalization and verification then take a further 17 hours, running on GPT-6 Astra.
OpenAI publishes the result and states in the same post that training of the model is ongoing and its performance continues to improve. As of this page's update, no name or date has followed.
Three tiers: confirmed, unpublished, and what this page will not print
Officially confirmed (quotable from the post)
1. An internal model exists that is significantly more capable than GPT-6 Astra. 2. It has been in training since 2026-08-28 and training is ongoing. 3. A multi-agent system powered by it produced the Navier-Stokes resolution on 2026-09-05. 4. Lean verification ran on GPT-6 Astra and took an additional 17 hours. 5. OpenAI does not intend to claim the Millennium Prize. All five can be found verbatim in the 2026-09-08 post.
Not published by OpenAI (do not plan around it)
Name, API model id, release date, price, context window, output limit, and whether it will be released publicly at all - none of these has been given. Several codenames circulate in the community, but they have no official source and this page repeats none of them: writing a community label as if it were an official model name is the most common and least reversible error in this topic. If you need one sentence for your team, the only honest one today is: OpenAI has confirmed it is in training and has given no date.
How it relates to GPT-6 Astra
GPT-6 Astra - callable today
Released 2026-09-03, API model id gpt-6-astra, 1,050,000 context and a 128,000 output cap that includes reasoning tokens, list price $10 per 1M input and $50 per 1M output. The 17 hours of Lean verification above ran on this model, so it is already capable of that class of work.
The internal model - out of reach
All OpenAI gave is the relative phrase "significantly more capable". No id means no way to call it; no date means no way to schedule around it. Its practical implication for you today is to write code where the model name is one line of config, rather than to wait.
The preparation that is actually worth doing
Lift the model name out of your code and into configuration or an environment variable, so switching costs one line. Every OpenAI model shipped this year (the three GPT-5.6 tiers, GPT-6 Astra) speaks the same Responses API; migration friction comes from parameters rather than from the interface. Astra drops temperature, top_p and top_logprobs, has no "none" reasoning tier so you start at "low", and requires tool calls to go through Responses. Concentrate those differences in one adapter layer and the next model - whatever it is called, whenever it lands - costs you a config change rather than a code change. This is work you can finish today, and it does not depend on anything unreleased.
How to prepare on QCode
One QCode key already calls gpt-6-astra and the three GPT-5.6 tiers; switching is a model id change, with no new endpoint and no new auth. That means you can build the adapter layer described above on Astra today and measure your cost baseline now. When the next model is actually released publicly, the work left for you is adding a model id. We will not promise a launch date for any unreleased model - OpenAI has not given one, so we certainly should not. Check the /models page for live status.
Frequently asked questions
What is OpenAI's next model called?
OpenAI has not published a name. The 2026-09-08 post refers to it only as an "internal model" throughout. Any specific name is unofficial, including the codenames circulating in the community and including "GPT-7", which is simply extrapolated from the numbering.
When will it be released?
No date has been given. The official wording is that training is ongoing and performance continues to improve. That statement by itself rules out any confident claim of an imminent release - a model still in training has no release schedule. Any source quoting a specific date is not OpenAI.
How much stronger is it than GPT-6 Astra?
OpenAI wrote only "significantly more capable". The chart in the post plots pass rate against test-time compute and carries no citable numbers, so this page gives no percentage. The one checkable capability data point is that a multi-agent system powered by it resolved Navier-Stokes.
Can I use it now?
No. OpenAI explicitly calls it an internal model and says training is ongoing. With no API model id there is no way to call it. The strongest OpenAI model you can call today is GPT-6 Astra, released 2026-09-03.
What does the Navier-Stokes result actually prove about its strength?
It shows the system can produce a Lean-verifiable proof for a problem open for roughly 90 years, which is hard evidence. Three caveats belong with it: that result came from on the order of 10,000 concurrent agents working about 88 hours, not from one conversation; the verification step ran on GPT-6 Astra; and OpenAI itself declines to claim the Millennium Prize and calls the result "a snapshot in time" rather than a culmination.
Should I wait for it, or move to GPT-6 Astra now?
Waiting for something with no date is not an executable plan. The better move is to lower your switching cost now: put the model id in config and concentrate parameter differences in one adapter layer. Then whenever the next model lands, switching costs one line. The usage patterns and cost baseline you build on Astra today stay valid either way.
Sources
Every quote and number on this page comes from OpenAI's post of 2026-09-08, "On the Navier-Stokes Millennium Prize Problem" (openai.com/index/navier-stokes-solution/), and from "An Alien Mind" of 2026-09-06 by Chief Scientist Jakub Pachocki (openai.com/index/an-alien-mind/). GPT-6 Astra's specification and pricing come from the official model documentation and pricing page. This page cites no secondary coverage and repeats no codename that lacks an official source. The update date appears in the page footer.
Lower your migration cost first, then wait for the new model
One key calls GPT-6 Astra and the three GPT-5.6 tiers; switching is a model id change. When the next model lands, your job is adding one line.
Related pages
Complete GPT-6 guide
GPT-6 Astra specification, pricing and practical usage.
GPT-6 Astra migration guide
The four parameter changes needed when moving from GPT-5.6 to Astra.
GPT-6 Astra quota burn
How Astra consumes quota in Codex, with measured community observations.
This page describes a model that has not been released. Every item under "officially confirmed" can be checked verbatim in the official posts listed above. Name, release date, price and API model id remain unpublished; this page makes no guesses about them and promises no launch date for any unreleased model.