OpenAI’s core product lead posted four words in French on X: Le GPT Fat. Thibault Sottiaux, a Belgian who studied applied mathematics at KU Leuven and now runs ChatGPT and the core product line, has a habit of writing this way before a release. His previous post of the kind, 6.1 coming soon on 4 October, was followed about three weeks later by GPT-6.1 Terra.
Then, in the early hours, two model identifiers that had never been published appeared in OpenAI’s API model list: rc-ada-or-vm and rc-ada-sh-vm. Minutes later they were gone.
Unlike three Aurora models that sat in the list for about 26 hours last month before anyone noticed, this pair surfaced and vanished within minutes, which looks more like an internal typo caught immediately. Developers fast enough to fire a request hit 503 errors reporting no available compute.
Why one developer’s post moves the market
Sottiaux has an unusual standing among developers. He answers threads personally and has been known to reset Codex quotas for people who ran out, which earned him the nickname cyber godfather. That hands-on style gives his posts their own reading. Every cryptic message gets treated as a pre-release signal.
So a French codeword plus two identifiers was enough to conclude that OpenAI is preparing something large. Fat is the industry’s shorthand for big, and that lines up with what OpenAI staff have said publicly.
An employee, Ted Sanders, explained the ordering of four models on Hacker News. Luna is the smallest, then Terra, then Sol, then Astra at the top. He wrote that the performance gap is broadly determined by a single variable, size, and that bigger models are more expensive and more reliable.
Price makes the hierarchy concrete. Astra charges USD 10 per million input tokens and USD 50 per million output tokens. Luna, the smallest, charges USD 0.1 and USD 0.5. A Fat version of Astra implies something heavier than the current flagship.
What 6.2 would mean
By Sanders’s account, a version number now carries three things at once: release order, kinship with other models of the same generation, and the size of the jump. 6.1 signalled a model after 5.6 that shares training data and character with the other 6.1 releases, with an improvement not large enough to justify the name 7. On that logic 6.2 is a smaller step than 6.1 was.
The cadence fits. Astra shipped in September, October brought 6.1 Sol and 6.1 Terra, and a 6.2 within two months matches a company whose release density keeps increasing.
The identifiers themselves carry contradictions. The rc prefix almost certainly means release candidate. The middle segment, ada, is the name OpenAI gave its earliest and smallest model. Pairing the smallest legacy codename with a suffix that emphasises Fat is either an engineer’s joke or a test carrying two opposite sizes in the same batch. The trailing or, sh and vm look like internal fragments distinguishing use cases.
What supports the reading is that shadow models keep happening at this company. On 17 October three Aurora-named models appeared quietly in the API list, in red, green and blue, and stayed for about 26 hours before third-party tracking accounts found them and they were deleted. Calls returned 503 then as well.
Earlier in the year three tape codenames appeared, maskingtape, gaffertape and packingtape. Seventeen days later OpenAI shipped GPT Image 2. In February an Aurora Alpha build surfaced on OpenRouter, and three days later GPT 5.3 Codex Spark was out.
A mysterious model appearing in the API shortly before a major release has become a fixture. Third-party accounts poll the model list continuously, so no new identifier escapes them, and the accidental leak has become part of the launch process.
Astra carries its own history
The Astra codename is not neutral. Earlier this year OpenAI slowed its release schedule after Astra reached a critical rating in an evaluation of autonomous cyber-attack capability, and put a set of controls around it: isolated test environments, encrypted weights, round-the-clock monitoring. The caution around the Astra family is well established, and a 6.2 carrying that name inherits heavier scrutiny.
A shadow model pulled within minutes is also consistent with a company that tests capacity internally before signalling externally. The more likely explanation is a capacity canary rather than a formal rehearsal. OpenAI has said nothing officially, and the two identifiers could equally be a configuration error.
What is not in doubt is the information asymmetry the incident reveals. A withdrawn API entry became a global developer guessing game within hours. OpenAI’s unpublished list now functions as the industry’s weather forecast, and whether or not GPT-6.2 Astra arrives this month, the shadow models have already told the players that there are more cards under the table.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.ofweek.com/ai/2026-11/ART-201718-8120-30710355.html.