Rumor Mill AI Minute Newsroom 2026-08-28

Rumour: OpenAI is said to have finished training a 10-trillion-parameter model called Bel — and every retelling traces back to one anonymous post

This is an unverified rumour and OpenAI has said nothing about it. On 25 August an account on X called @synthwavedd claimed the company had completed a pretraining run codenamed Bel, successor to one codenamed Doug, at more than 10 trillion total parameters, and that Bel is meant to become a base model beyond GPT-6. The post passed 660,000 views and was picked up by several technology sites within a day. Every one of those write-ups traces back to that single post, and several of them garble it by swapping the two codenames. Nothing here has a benchmark, a system card, or a statement from OpenAI attached to it. Worth noting alongside the claim: the reinforcement-learning work the rumour says still has to be done is the same kind of work OpenAI publicly paused on its Astra model this month.

Why it mattersWe are publishing this as a rumour, not as news, because it shows how the biggest claims in AI now travel. A single anonymous post about a model nobody outside one company can see became a day of headlines, and the number in it — 10 trillion — is exactly the sort of figure that cannot be checked from the outside even after a model ships. Treat parameter counts from unnamed sources as marketing until a lab puts its name to them. If OpenAI confirms any of this, we will report it as a real story and add a note here.

⚠ Unverified rumor — treat with caution

Read in the app — free, in 9 languages

Rumour: OpenAI is said to have finished training a 10-trillion-parameter model called Bel — and every retelling traces back to one anonymous post

Related stories