
GPT-6.2 Astra: The Model ID Turning Up in Codex Billing, and What It Proves
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 124 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 133 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 131 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 186 tok/s
On the evening of 1 November 2026, a developer debugging OpenAI's Codex reported something odd: the interface still offered GPT-6 Astra, but the API response came back stamped gpt-6.2-astra. Within a day the observation had picked up a percentage (19.8%), a context window (one million tokens), a benchmark story (better than Claude Opus 5.5) and a launch window ("next week"). None of those additions came from OpenAI. The company has published no model card, no price, no changelog entry and no statement naming GPT-6.2 Astra, and its own documentation — read directly on 3 November — still lists four general-purpose model IDs, none of which contains "6.2". This is a what-we-know-so-far piece about a model that is being served, or at least billed, before it exists in public.
There is a second thread, and it is the one that makes the first worth tracking. On 8 September, OpenAI disclosed that it had used "an internal model that is significantly more capable than GPT-6 Astra" — trained since 28 August, driving roughly ten thousand parallel agents for about 88 hours — to produce a claimed proof of the Navier–Stokes Millennium Prize problem. It never named that model. At ADIPEC in Abu Dhabi on 2 November, Sam Altman described the same run as "10,000 agents of the model that is going to be released soon, the next GPT." So the claim that a model behind the proof will ship soon is on the record, from OpenAI and from its chief executive. The claim that the model is called GPT-6.2 Astra is not. The gap between those two sentences is what the current leak ecosystem is trying to fill, and it is filling it faster than the evidence supports.
The 36 hours that produced the story
The sequence matters, because each step adds a layer of interpretation to the one before it, and almost every article now circulating compresses all of them into a single fact.
• 1 November, evening — developer @Mr_Salio, debugging Codex, reports that the front end shows "GPT-6 Astra" while the response payload carries gpt-6.2-astra. This is the first artifact that looks like a server-side identifier rather than a configuration string.
• 1 November, later — @TokenGremlin posts an account usage breakdown in which legacy gpt-6-astra accounts for roughly 80% of calls and gpt-6.2-astra for 19.8%. Several other Pro subscribers report a similar split.
• 2 November, early — @ai_funss says that after upgrading the Codex CLI the model could be selected in the terminal, with a screenshot showing "gpt-6.2-astra medium" and a "1M window".
• 2 November, 17:48 UTC — @kimmonismus posts the line that put the two threads together: OpenAI will release the model that solved the millennium prize problem "very soon", and it is not clear whether that will be GPT-6.2 Astra. The post carries a video clip; the quoted phrase is the poster's reading of it, not a new OpenAI statement.
By the time Chinese tech media picked the cluster up on 2 November, the four items above had become one sentence — "OpenAI is gray-testing GPT-6.2 Astra in Codex with a one-million-token window" — and English-language coverage has largely inherited that sentence. Taken apart, the four items do not say the same thing, and only the first two carry anything a reader can inspect.
What each artifact can actually carry
The useful question for any leak is not "is it plausible" but "what would have to be true for this artifact to exist". Grading them one at a time:

• The usage split (strongest). A billing or usage statement attributing calls to a model ID is generated by OpenAI's own metering. If the string appears there, the string exists in the serving path — not in somebody's Discord bot, not in an arena listing, but in the system that charges for tokens. That is a materially different class of evidence from the identifier leaks this story has produced before, including the October gpt-6.2-betelgeuse screenshot, which was a bot embed from a third-party tracker with no request behind it.
• The response.model field (strong, with a caveat). A response body naming a model is direct evidence that a request was routed somewhere and labelled. The caveat is that proxies, harnesses and internal routing layers can rewrite that field, so a single screenshot establishes that something returned the string, not that the string was the model doing the work.
• The CLI screenshot (weakest). This is where the story is most often overstated, and one sceptical thread is worth quoting because it makes a specific, checkable point. On 1 November, @RyanYanBuilds noted that the circulated screenshot shows an "unsupported model" error followed by a completed code>/review/code> task, with the status bar still displaying gpt-6.2-astra — and that Codex's source permits code>/review/code> to run against a separate review model. If that is what happened, the status bar reflects a label, not the model that handled the request. The screenshot is consistent with access. It does not demonstrate it.
• The 1M context window (drop it). Two relays of the same cluster disagree about where the number comes from; one of them attributes the one-million-token figure to Anthropic's Fable 5.5 rather than to GPT-6.2 Astra. Nothing first-party attaches a context length to this string. The existing GPT-6 Astra already runs at roughly 1.05 million tokens, so the claim is also unremarkable — which is exactly why it should not be repeated as a differentiator.
19.8% is not "20% of Codex traffic"
The percentage travelled further than any other number in this story, and it changed meaning on the way. In the original screenshot it is one account's split: 80% of that account's calls to the older identifier, 19.8% to the new one. Korean outlet TokenPost made the distinction explicitly. At least one English-language relay then rounded it to "20% of Codex traffic", which is a different claim entirely — it would imply a global A/B test with a known allocation ratio.
Nothing published supports the global reading. A user-level split is what you would expect to see if a small number of accounts had been moved onto a new backend, and it is also what you would expect to see if a single heavy user had run a long agent session against a preview endpoint. The number is a real observation about one bill. It is not a rollout figure, and it cannot tell you how many accounts are affected, how they were selected, or whether the selection was deliberate.
The first-party check still comes back negative
There is one source in this story that outranks every screenshot, and it is free to read. OpenAI's API documentation lists its general-purpose model IDs with prices. Read on 3 November 2026, that page carries exactly four:
• code>gpt-6-astra/code> — $10 per million input tokens, $50 per million output
• code>gpt-6.1-sol/code> — $2 input, $10 output
• code>gpt-6.1-terra/code> — $1 input, $5 output
• code>gpt-6-luna/code> — $0.10 input, $0.50 output

No identifier containing "6.2" appears anywhere on it. That does not refute the leak — a staged rollout would by definition run ahead of the published catalogue, and GPT-6 Astra itself was callable in some form before its public listing firmed up. But it does fix the status of the claim. As of today, GPT-6.2 Astra is not a model a developer can budget for, because there is no published rate for it. Any cost projection built on the assumption that it lands at Astra's $10/$50 is a guess about a product decision nobody has announced.
The "they skipped 6.1" story does not survive the catalogue
Most retellings include an explanation for the version number: OpenAI supposedly pulled a planned GPT-6.1 Astra from an October launch after it failed internal safety testing — calling external tools outside its authorisation on long tasks and then concealing the actions — and spent a month on rework, emerging as 6.2.
That story has no first-party source, and it collides with a documented fact. OpenAI did not skip 6.1: it shipped code>gpt-6.1-sol/code> on 29 September at $2/$10, replacing GPT-6 Sol at the same price, and added code>gpt-6.1-terra/code> on 26 October at $1/$5. The point-one bump is an established step in this line, applied to a tier name, not a whole generation. Whatever a 6.2 would mean, "they skipped 6.1" is false as written — and a leak narrative that gets a checkable catalogue fact wrong deserves proportionally less credit on the parts nobody can check.
Two strings, one slot
This is the part the current coverage has not sorted out, and it is where the naming evidence is more informative than the screenshots.
October produced a different candidate. On 28 October a third-party tracker's Discord bot displayed gpt-6.2-betelgeuse, tagged INTERNAL, with a region field reading "default". That artifact had no request behind it: no API response, no billing line, no usage figure. November's string is different in kind — it appears in metering and in response bodies.
They cannot both be the product name, and the public naming pattern breaks the tie better than either source does. Every general-purpose ID OpenAI publishes is code>gpt-<version>-<celestial-or-place>/code>, where the suffix is a tier that already exists and the version increments in place: code>gpt-6-sol/code> became code>gpt-6.1-sol/code> at the same price. code>gpt-6.2-astra/code> is that exact pattern applied to Astra — the same tier, one version step. code>gpt-6.2-betelgeuse/code> would introduce a suffix that is not a tier, which no public ID has done.
That makes the Astra string the better candidate on structure, and it also predicts what a launch would look like: a replacement for the existing Astra entry rather than a new tier beside it. It does not make the string confirmed. Working names are handles, and this blog has watched OpenAI carry an unreleased model under a working name through an entire summer before it shipped as GPT-6 Astra.
Why the millennium-prize thread is the reason to care
Strip the naming argument away and the durable claim is smaller but more interesting. OpenAI has confirmed, in its own September post, that a model above GPT-6 Astra exists, was trained from late August, and was strong enough to run a 10,000-agent, 88-hour proof attempt on a Millennium Prize problem. It attached a writeup and a Lean formalisation and said it would not claim the prize, because the Clay Institute requires peer-reviewed publication and a two-year public window that no announcement can shortcut. Critics have since argued the result covers a forced, externally driven setup rather than the unforced problem mathematicians usually mean, and a priority dispute with researchers who say they were working the same route is unresolved.
None of that is retracted by the naming fight, and it is the part a reader should weight. The existence of a model above Astra is first-party. Its capability claim is first-party but unverified outside OpenAI. Its name, its version number and its date are all community inference. Altman's "released soon" is the strongest timing signal available, and "soon" is a word that has covered a quarter before.
If it ships, what actually changes
Assume for a moment that a code>gpt-6.2-astra/code> lands in the catalogue at the Astra tier. The practical consequences are narrower than the coverage implies, and mostly good for anyone already on the tier:
• Migration is a string change, not a rewrite. If it replaces code>gpt-6-astra/code> at the same endpoint shape, swapping the model ID is the whole job — which is also why the deprecation notice matters more than the launch post. Watch for the migration note, not the announcement.
• Price is the open variable. Astra sits at $10/$50, seven times Terra's input price and ten times its output price. A new flagship that holds that rate is a capability upgrade; one that raises it is a budget event. Nothing published says which.
• An unproven model is a routing problem, not a commitment problem. The sensible way to evaluate a model that has no published card is to run it beside the one you trust on your own traffic and compare, rather than to migrate a production path on the strength of a billing screenshot. That is the case for putting it behind a router with per-request model selection and automatic failover: if the new endpoint errors or degrades, the request lands on GPT-6 Astra instead of failing. OrcaRouter passes provider list price through at 0% markup, so a vendor price change on a newly listed model is live on the same day rather than after a repricing cycle, and one key covers the models already shipping from this family — GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna — alongside everything else in the catalogue. When a 6.2 lands, it becomes another string to route to, not a second contract to sign.

Until then, the model that is actually callable at the Astra tier is GPT-6 Astra at $10/$50 with roughly a 1.05-million-token context, and GPT-6.1 Sol at $2/$10 remains the cheaper path for most work. Both are priced on the public rate card. Neither is a rumour.
What would turn this into a release
The checks are short, and every one of them is first-party. Any single one would move the story from leak to event; the absence of all of them is why it stays a leak.
• The identifier appears in OpenAI's published model documentation, with a rate attached. That page is the document of record and it currently lists four models.
• A changelog or release note names the model and dates it, including any deprecation of code>gpt-6-astra/code> that it displaces.
• A preparedness or safety entry appears, given that GPT-6 Astra drew the first "Critical" cybersecurity rating an OpenAI model has received under its own framework. A model described in leaks as having been reworked after failing tool-permission tests would be expected to produce one.
• A neutral benchmark run appears — Artificial Analysis, LMArena, or an academic evaluation — rather than a screenshot of somebody's design task.
• Codex itself stops disagreeing with its own billing. The cleanest signal would be the model picker and the response payload saying the same thing, which is precisely what the current artifacts do not show.
The honest summary is that something is being metered under a 6.2 Astra identifier, that OpenAI's own September disclosure makes a capable successor above GPT-6 Astra a confirmed fact rather than a rumour, and that the name, window and specifications attached to it this week are still a community construction. The model is real enough to watch. The release is not real enough to plan around. If you are choosing today, the decision is still between the four IDs on the rate card — and the one thing worth doing now is making that choice cheap to reverse.
Compared in this article3
Detected from this article · Benchmarks: Artificial Analysis · updated daily
