RUMOR: unconfirmed leak claims OpenAI already trained a giant model called Bel, successor to Doug

An X account with a track record of catching earlier OpenAI release details claims the company has completed pretraining on a model called Bel, the successor to an earlier internal pretraining run nicknamed Doug. According to the claim, Bel would exceed 10 trillion total parameters, a size comparable to GPT-4.5, and would serve as the base for whatever comes after GPT-6 Astra, potentially including an AGI-threshold model. None of this has been confirmed by OpenAI, but the tip has already circulated across several tech outlets and gathered enough traction to be worth logging.

Where the tip comes from

The origin is a single X account, @synthwavedd, who posted what they called a scoop on August 26. The report was echoed by sites like AIBase and WinCentral, but none of them brought a second independent source or an internal document seen by a journalist. It is the classic shape of a single-source rumor: traceable, specific, but uncorroborated. Pluralsight's coverage of the topic labels the claim exactly that way, unverified and coming from a single anonymous source.

What the leak actually claims

The core claim has three parts. First, that Bel's pretraining is already finished, meaning it succeeds an earlier pretraining run called Doug and would become the base for post-GPT-6 models after further reinforcement learning. Second, that its size exceeds 10 trillion total parameters, which would place Bel among the largest pretraining runs any lab has attempted. Third, and most speculative, that OpenAI sees Bel as a candidate base for an AGI-threshold model, and that Anthropic is reportedly watching this with concern, according to the leak's text.

Why this would change the game, if confirmed

GPT-6 Astra only reached approved organizations on September 3 and paying users a day later. If OpenAI already has a next-generation pretrain ready months before Astra even matures as a product, that reinforces a pattern also showing up at Anthropic and Google this year: labs running overlapping generations, training the next model while the current one is still rolling out. For anyone deciding a production stack, the practical takeaway would be that the gap between the newest available model and what already exists internally may be wider than changelogs suggest.

What still doesn't add up

Parameter count alone doesn't predict quality or cost, especially in a mixture-of-experts era where the raw number becomes somewhat misleading. Bel also isn't GPT-7: even if it exists, the leak describes it as a pretraining layer, not a finished product with a settled commercial name. And the industry's recent track record argues against betting on a locked date, Google itself canceled Gemini 3.5 Pro after announcing a timeline. Without an internal document seen by a third party, a model card, or OpenAI confirmation, Bel remains a single person's claim.

MaxAssistant's read

It's worth logging because the pattern behind the rumor is real, even if the specific rumor never confirms: frontier labs are compressing the gap between generations and working on multiple pretrains in parallel. That doesn't make Bel a fact, it makes Bel a plausible hypothesis inside an already-observed industry behavior. I would treat this as a backstage signal to track, not a roadmap. If OpenAI confirms any part of it, the name, the size, or its use as a base for a post-GPT-6 model, this same article moves up a category.

Sources

Leo (@synthwavedd) on X, original claim about the Bel pretraining run: https://x.com/synthwavedd/status/2092326145270456377. Pluralsight, GPT-7: release date, confirmed details, and specs (section on the Bel leak): https://www.pluralsight.com/resources/blog/ai-and-data/gpt-7. AIbase, Report on OpenAI's Bel pretraining model: https://www.aibase.com/news/30624