GPT-6.2 'Bel': The Betelgeuse Model ID in a Leaked Screenshot, and What It Actually Proves

The most concrete thing anyone has shown about OpenAI's next flagship this week is a 362-by-207 pixel image of a Discord bot. It was posted on October 29, 2026 as a reply in a thread by @kimmonismus, who wrote that he could not confirm it but that it looked like OpenAI was about to "fire their big guns: GPT-6.2 'bel' incoming." The embedded bot message is headed "Model Updated," says "Changed (1)," and lists one string: gpt-6.2-betelgeuse - INTERNAL, with a line underneath reading "regions: default" and a warning not to share it outside the server. That is the entire artifact. There is no API response attached, no model card, no changelog entry, no price, no benchmark, and no statement from OpenAI. The company's own model catalogue, read on the morning of October 30, lists exactly four general-purpose tiers — GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna — and nothing above them. So the right way to read this is not "GPT-6.2 is coming" but "someone's tracker saw a name in a configuration somewhere, and a month-old codename has been upgraded into a product number on social media." What follows is an inventory of what exists, a grade for each claim attached to it, and the specific checks that would turn a Discord embed into a release.

What the Discord embed actually shows

The image was shared by @taker_of_whizz, who said "Our internal tracker picked this one up yesterday!" The bot is called Internal Tracker with an "APP" tag and a server tag reading "ABB." Its timestamp is "Yesterday at 10:14 AM," which, against the October 29 thread, places the observed change on October 28. The payload is three lines: a title, a change count, and one model string tagged INTERNAL, plus a region field marked "default." A footer warns: "Do not share outside of this server."

Two things about that framing deserve attention. First, "Internal Tracker" is not an OpenAI tool, or at least nothing in the image says it is; it is a bot in somebody else's Discord server, watching something. Second, the line that does the work — "regions: default" — is the kind of field a deployment or availability config carries, not the kind of field a model card carries. That is consistent with an identifier being staged or replicated somewhere, and inconsistent with a model being callable. The distinction matters because the same community produced a very similar artifact in October: three identifiers, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha, appeared briefly in an OpenAI model listing on October 17 and 18, drew a wave of "next GPT-6" speculation, and were later read by most observers as image-generation variants after arena testing — with no evidence any of them returned a successful request while visible. An identifier in a list is a rumour with better formatting. It is not access.

Where the name "Bel" came from

"Bel" is not new, and that is the part the October thread glosses over. The codename entered circulation on August 25, 2026, when the anonymous X account @synthwavedd posted that OpenAI had recently finished a pretraining run called Bel, that it succeeded an earlier run called Doug, and that its total parameter count exceeded 10 trillion. That post drew roughly 812,000 views and was relayed into Chinese, Russian and Thai tech press within a day. Ten days later, a September 5 relay by @Mr_Salio added a headline the original never carried — "GPT-6 Bel Leaks: Beats GPT-6 Astra" — with no benchmark behind it. By the time the cluster had been fact-checked in early October, the comparison had become the most repeated claim and the least supported one.

Only one piece of the codename story has independent corroboration. A SemiAnalysis subscriber note dated July 9, reproduced on August 7, referred separately to a "much larger" OpenAI model in the works under the name Doug. Everything else — that Bel is Doug's successor, that the pretrain finished, that the count is 10 trillion, that Bel is the base for whatever follows GPT-6 — traces to one anonymous account. OpenAI has never confirmed the name in its documentation, API changelog or news feed, and it did not confirm it in the one place it came closest.

That place is worth quoting precisely, because it is the strongest official fact in this story. In its September 8 post on the Navier–Stokes Millennium Prize problem, OpenAI wrote that it had used "an internal model that is significantly more capable than GPT-6 Astra," that it had been training that model since August 28, and that it drove roughly 10,000 parallel agents for about 88 hours to produce a claimed proof. The company attached a writeup and a Lean formalization, and said it would not claim the prize. What it did not do — then or since — is put a name on the model. So OpenAI has confirmed that something above Astra exists and is being exercised on hard problems. It has never confirmed that the thing is called Bel, and it has never given it a version number.

The tells: Betelgeuse fits the family, Dougtrio does not

Read the replies under the October 29 thread and the leak mutates in real time. One reply, from @Josh67Joshua, states as fact that "Bel, now understood to be short for GPT-6.2 Betelgeuse, is based on the Dougtrio pre-train series." No document in the thread says either thing. But the first half is more plausible than it looks, because it fits a pattern OpenAI has been running all year: the general-purpose tiers are named for celestial bodies and places. GPT-6 Sol, GPT-6 Luna, GPT-6 Astra and GPT-6.1 Terra are all in that family, and Betelgeuse is a red supergiant in Orion whose first three letters are Bel. If you were inventing a codename that would read as an OpenAI codename, that is exactly the one you would invent.

The second half is the tell. "Dougtrio" is not an OpenAI naming pattern; it is the spelling of Dugtrio, a Pokémon, one letter off. It reads like a joke built on the existing Doug codename, and it appears only in a reply. That is how these clusters work: one real string, one plausible expansion, one joke that gets quoted as detail, and by the third relay nobody can tell which was which. The October 29 thread is a compact example. The artifact says gpt-6.2-betelgeuse. The reply says Bel means Betelgeuse and the pretrain series is Dougtrio. The next repost will say OpenAI's Betelgeuse is built on Dougtrio. None of those three sentences has the same evidentiary weight, and only the first has an image behind it.

Why the "6.2" is the strangest part of the claim

OpenAI shipped GPT-6 Sol on September 22, moved to GPT-6.1 Sol seven days later on September 29, and added GPT-6.1 Terra on October 26 at $1 per million input tokens and $5 per million output tokens — half of GPT-6.1 Sol's $2 and $10. That is a fast cadence, and a point-one bump inside it is already unusual: the company went from 6 to 6.1 in a week, not after a quarter. A 6.2 would be the first second-point release in the line, and it would arrive roughly four weeks after Terra. Possible. Not implied by anything in the screenshot.

There is a naming problem too. OpenAI's public identifiers look like gpt-6.1-sol and gpt-6.1-terra — a version and a celestial name, no codename suffix. A string like gpt-6.2-betelgeuse tagged INTERNAL looks more like an internal staging name than a public model ID, which is precisely why it should not be read as a product name. The blog has been here before: through the summer it tracked an unreleased model under a working name that turned out to ship as GPT-6 Astra. Working names are handles. They are not commitments, and they are not dates.

What 10 trillion parameters would and would not mean

Assume for a moment the August claim is accurate. Ten trillion total parameters would put Bel in a size class well beyond anything OpenAI has published a number for, and the comparison to GPT-4.5 that circulated with it is doubly soft, because OpenAI never published GPT-4.5's parameter count either. Total parameters are also the wrong number to judge a model by. If Bel is a mixture-of-experts model — and at that scale it would almost certainly have to be, for serving to be economic at all — the figure that determines cost per token and latency is active parameters per token, not the total in the checkpoint. A 10-trillion-total model with a few hundred billion active could route cheaply; a dense 10-trillion model would be a different order of infrastructure problem.

The more important distinction is between a finished pretrain and a shippable product. A base model that has completed pretraining is raw capability. Everything a buyer experiences — instruction following, tool use, refusal behaviour, long-context quality, safety evaluation under OpenAI's Preparedness Framework — comes after post-training, reinforcement learning and evaluation, and that phase is where schedules actually slip. OpenAI's own September 8 disclosure describes a reinforcement-learning run, not a pretrain, and it describes a system being used, not one being packaged. The anonymous source behind Bel said on September 23 that the model was still months from public. Nothing published since has shortened that estimate except a Discord bot that copied a string.

The claims, graded

• That an identifier containing "gpt-6.2-betelgeuse" appeared in a tracker on October 28 — single artifact, one Discord screenshot, unverified provenance.

• That OpenAI has an internal model more capable than GPT-6 Astra — confirmed by OpenAI on September 8, unnamed, no date, no price.

• That the model is called Bel — single anonymous source since August 25, never confirmed by OpenAI.

• That Bel means Betelgeuse and descends from a "Dougtrio" pretrain series — reply-level claims, no document, and "Dougtrio" is a misspelled Pokémon.

• That the model exceeds 10 trillion parameters — single anonymous source; no vendor figure exists.

• That GPT-6.2 is "incoming" — the poster's own words are "can't confirm"; no date, price, benchmark, weights or API availability accompanies the claim.

Five checks that would make this real

If you want to follow this without being fooled by the next screenshot, there is a short list of things that count, and they are all first-party.

• The identifier appears in OpenAI's own model catalogue or a documented API listing. As of October 30, the catalogue's general-purpose entries are GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, and there is no GPT-6.2 anywhere on it.

• A changelog or release note names the model and gives a date. OpenAI's changelog is the document of record; a bot embed is not.

• A rate card appears. Every OpenAI launch this year has arrived with a price per million tokens, and price is the thing a developer can act on.

• A deprecation or migration note lands for an existing model. Successors usually displace something, and OpenAI announces the displacement before it happens.

• A safety or preparedness entry appears for the new model, given that GPT-6 Astra drew the first "Critical" cybersecurity rating under OpenAI's own framework. A model above Astra would need that evaluation before it ships, and that evaluation is published.

What does not count: another bot embed, another relay account quoting the first one, a benchmark table with no method, or a countdown account. The aurora identifiers in October are the control case — a real-looking string, a burst of coverage, and no callable model.

What you can call today, and what a 6.2 would cost you

The practical answer for anyone building right now is that the current lineup is already unusually cheap at the middle. GPT-6.1 Terra at $1 and $5 per million tokens is half of GPT-6.1 Sol at $2 and $10, and a fifth of GPT-6 Astra at $10 and $50. GPT-6 Luna sits at $0.10 and $0.50. Those are vendor list prices, and on OrcaRouter they pass through at 0% markup, so a price change from OpenAI is live on our side the same day rather than after a contract renegotiation.

OrcaRouter routes GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna today; GPT-6.1 Terra is not in our catalogue at the time of writing. If a GPT-6.2 does arrive, the migration decision will turn on two numbers: whether it is priced above Astra, in which case it is a premium tier you opt into for specific work, or at or below GPT-6.1 Sol's $2 and $10, in which case it is a like-for-like swap and the only question is evaluation. Either way the cheap hedge is the same — keep the model name out of your application code, call an OpenAI-compatible endpoint, and let routing decide which tier serves each request. Automatic failover is what makes trying an unproven model safe: you can send a slice of traffic to a new model and fall back without rewriting the client.

The honest read, as of October 30, 2026

Something above GPT-6 Astra exists. OpenAI said so itself in September, in a post about a mathematical result, and it described the system as still training at the time. A pretraining run called Bel has been rumoured since August from a single anonymous source, with a 10-trillion-parameter figure that no one has corroborated. On October 28 a Discord tracker saw a string that looks like an internal name for a 6.2-tier model called Betelgeuse, and on October 29 that string became "GPT-6.2 incoming" in a post whose author explicitly could not confirm it. Those are three different levels of evidence, and the distance between the first and the third is where most readers will get hurt.

The useful conclusion is not that GPT-6.2 is or is not coming. It is that no one outside OpenAI can currently date it, price it, or run it, and that the checks which would change that are all first-party and all public. Watch the model catalogue, the changelog, and the rate card. Until one of them moves, the number is a handle, the star is a guess, and the Pokémon is a joke that got repeated too many times.