CellCog traced the claim of more than 10 trillion parameters, or values a model learns during training, to an August 25 post about an OpenAI training run called Bel. An October 29 screenshot reproduced by OrcaRouter showed a name containing GPT-6.2. OpenAI has published no statement connecting the size claim and the name.
From an August claim to an October screenshot
In August, the X account @synthwavedd claimed that OpenAI had finished the first stage of training for a model it called Bel. The account put Bel at more than 10 trillion parameters and said it followed an earlier effort called Doug. SemiAnalysis had published part of a July note that described Doug as a larger OpenAI model. The excerpt gave no size and did not mention Bel.
In October, the X account @taker_of_whizz posted an image that appears to show a Discord bot flagging a change to an entry named gpt-6.2-betelgeuse and marked internal. The image gives no source for that name. SnowRock could read published accounts of both posts. The original X posts were unavailable during the review.
The research model OpenAI did describe
In September, OpenAI said it used an unnamed internal model to work on a long-standing problem about the equations that describe fluid motion. OpenAI described the model as significantly more capable than GPT-6 Astra. It said researchers started with an existing model and trained it further by rewarding better results. OpenAI gave no parameter count for the internal model or the one it started with.
The August claim refers to a model’s total size. Some models use different groups of parameters for different inputs. Thinking Machines Lab says its Inkling 2 Small model, for example, has 594 billion parameters in all and uses about 17 billion at a time. Neither post gives a count for the parameters that would be used to answer a request.
