Codex users report GPT-6.2 Astra in their usage. OpenAI hasn’t announced it. What’s known

What people are seeing

By 07:37 UTC on November 1, 2026, Codex users were posting screenshots naming a model OpenAI hasn’t announced. @leo114119 wrote, in our translation, that although Codex shows GPT-6 Astra, the usage view shows GPT-6.2 Astra. Twenty minutes later, @Mashiro_0619 posted three more usage views, “collected from three users” by our translation, and called it a limited test. At 10:30 UTC, @Mr_Salio wrote: “The UI still GPT-6 Astra but the API response breakdown now shows ‘gpt-6.2-astra’.”

We read all five images. Four show the by-model view of ChatGPT’s usage history, with gpt-6.2-astra listed beside gpt-6-astra:

• In two seven-day views, it’s 1.6% and 19.8% of usage.
• In two views of October 31 alone, it’s 37.7% and 100%.
• Wherever the chart marks it, the usage falls on one day, October 31.

The fifth image is a four-line snippet of a response.completed event naming the model, with no account, time or request ID. TokenPost reports that other users said the name doesn’t appear in their records. We haven’t checked any of this against an account, and screenshots can be edited. What several people agree on is the label, and nothing more.

What OpenAI has said

Nothing about GPT-6.2. We checked its pages on November 1:

• The model release notes and API changelog both end with GPT-6 Rosalind on October 30.
• The API models page has no 6.2 model, and Codex’s model catalog on GitHub has gpt-6-astra, gpt-6.1-sol and gpt-6.1-terra, but no 6.2.
• The newest Codex CLI, 0.166.0 of October 29, adds Terra and doesn’t mention 6.2.

Three things OpenAI has said explain why people are primed for it, though none of them names GPT-6.2.

The 28 days. On October 4, OpenAI’s Thibault Sottiaux, who leads Codex, pledged: “Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset.” Trackers count November 1 as the last day. Early on October 31 UTC, Sottiaux wrote: “Next two days will be resets… banked… do with this what you may.” The next post read “Banked 1/2”. We hadn’t seen a second when we checked. The pledge promises an improvement or a reset, not a model.

The stronger internal model. On September 8, OpenAI’s Navier–Stokes post said OpenAI “used an internal model that is significantly more capable than GPT-6 Astra,” adding: “Since August 28 we have been training a new internal model.” It gave no name, date or plan to release it.

The Astra it scrapped. On September 28, OpenAI said it wouldn’t release GPT-6.1 Astra, which, in the words of its head of safety systems to CNN, “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done.” Quartz, citing The Wall Street Journal, says OpenAI intends to take its underlying model through more reinforcement learning “to build out subsequent entries in the GPT-6 family.” So some posters guess 6.2 is that model, reworked. That’s a guess: OpenAI hasn’t said so.

Three names, three levels of evidence

Name	Where it comes from	What backs it	From OpenAI
gpt-6.2-astra	Usage screenshots posted by X users, November 1	Four usage views and a snippet; other users don’t see it	Nothing
gpt-6.2-betelgeuse	A Discord bot called “Internal Tracker,” posted by @taker_of_whizz, October 29	One screenshot	Nothing
Bel	A post by @synthwavedd, August 25, calling it OpenAI’s next pretraining run, with more than 10 trillion parameters	No named source; repeated widely	Nothing

Reports call this an internal tracker, which reads as OpenAI’s. The poster calls it “our internal tracker” and doesn’t say where its data comes from. RuntimeWire reports that unnamed sources identify Bel as Betelgeuse. @ChrisGPT, an account that says it reported on Bel earlier and later heard of a new codename, is more careful: “Could this be the name change I’d heard about? Possibly. But I have not confirmed yet.”

The most repeated numbers come from an October 30 post by @Policifyai: that GPT-6.2 Bel would launch on the last day of the 28, score 72 on Artificial Analysis, have 12 trillion parameters, and cost $20 in and $100 out per million tokens. It credits an unnamed insider for the score, and what can be checked doesn’t help it:

• The score. Artificial Analysis’s leaderboard has no GPT-6.2 row and tops out at 59, with Gemini 4 Carbon. The chart in the post carries an X user’s handle, and it labels the 58-point bar Claude Opus 4.5, where Artificial Analysis lists Claude Opus 5.5.
• The price. $20 and $100 are what GPT-6 Astra already costs in Fast mode on OpenAI’s pricing page, so the figure tells you nothing new.
• The size. “Reportedly,” with no source. The August post said only more than 10 trillion.

When it might arrive

OpenAI hasn’t given a date, and no one citing one points to anything from OpenAI. WinCentral says “as early as next week,” as its own commentary. Polymarket asks when the next Astra numbered 6.1 or higher reaches the general public. When we read it on November 1, it had 44% on November 2 to 8, 15% on the week after, 17% on the week after that, and 8% on no release by January 31, with about $59,000 traded. That’s what a small group of bettors think, not information.

What to do

1. Look at your own usage, if you’re curious. In the desktop app, open Settings > Usage & billing. On the web, open your usage settings. The screenshots show an Analytics tab with a usage history and a by-model view, which we’ve translated from Chinese-language screens, so your labels may differ.
2. Don’t buy or change a plan over it. There’s no price and no list of plans. Two changes are confirmed: OpenAI doubled Pro 100 and Pro 500 usage on October 29, and ChatGPT and Codex share one allowance from December.
3. Hold your banked resets, if you can afford to. A banked reset is “saved to your account for later use,” and you “choose when to use it before it expires,” says OpenAI’s help article. Check the expiration in your usage settings. Waiting costs little, but Sottiaux’s hint is all there is to wait on, so use one if you’re blocked.
4. Keep your approvals and sandbox as they are. If the line is in your usage, you may have run something with no published model page or system card. The Astra OpenAI scrapped in September went out of scope in testing. That’s no finding about this one, but it’s no moment to loosen anything.

If GPT-6.2 reaches Codex

Enjoy runs Codex on your ChatGPT plan and asks Codex for its current models each time you open the settings, so a new model appears there without an Enjoy update, once your Codex offers it. If it’s missing, update Codex. If a new model burns through your allowance faster, you can switch the conversation to another agent, such as Claude Code, and it gets the messages it missed. Enjoy doesn’t add usage to a ChatGPT plan, and it can’t give you a model your account doesn’t have.

What OpenAI hasn’t said

• What gpt-6.2-astra is. It could be a limited test of a new model, or a label for something else.
• Whether it’s billed like GPT-6 Astra, which costs 250 credits per million input tokens and 1,250 per million output in Codex.
• Whether it’s the reworked GPT-6.1 Astra, or related to Bel, Betelgeuse or the Navier–Stokes model.
• Whether users are told when a request runs on something other than the model the picker names.
• Any price, plan, date or system card for a GPT-6.2.

We’ll update this memo when OpenAI fills these in.