OpenAI has completed pre-training a model with more than 10 trillion parameters, according to a leak that surfaced days before chief executive Sam Altman backed calls to slow the pace of AI development.
The claim, on which OpenAI has made no statement, centres on a base model codenamed Bel, said to be the successor to an earlier pre-training run called Doug.
Parameters are the internal settings a model adjusts during training, and a count above 10 trillion would place Bel in the same size class as GPT-4.5 and an order of magnitude above estimates for GPT-4.
OpenAI has published no parameter counts since GPT-3, so the figure rests on the leaker's word.
The claims trace to a post on X on 25 August by an account called leo (@synthwavedd) and spread through the tech press.
According to the leak, Bel is a raw foundation rather than a finished product and needs a round of reinforcement learning, the technique that refines a model by rewarding its better answers, before it reaches ChatGPT.
Several accounts describe GPT-7 as a possible snapshot of Bel's state at a chosen point in time.
Learning at two speeds
The leak's most ambitious claim concerns how Bel learns.
A fast weight layer is said to absorb lessons from verified proofs, code tests, experiments and tool traces while the model works.
A slower loop evaluates those lessons, fixes the ones that hold up into persistent weights and training recipes, and tunes the next round of learning.
The design would let the model change after launch rather than stay frozen at release, and distil the results into smaller models for everyday use.
The description echoes an academic framework called Fast-Slow Training, published in May, which splits learning into slow weights and fast, feedback-driven context.
Early accounts claim Bel outperforms Astra, OpenAI's newest release, on coding tasks.
Two generations ahead
GPT-6 Astra reached approved organisations on 3 September and paying ChatGPT users the following day.
If the leak holds, OpenAI would hold the skeleton of a further generation in house while customers get to grips with Astra.
The same leak claims OpenAI's internal assessment holds that Anthropic, its main rival, lacks the compute to answer its next generation, a claim that lands weeks before Anthropic's planned stock market listing.
Sceptics point to inconsistencies: sources differ on the order of Doug and Bel, and the "AGI threshold" label attached to the model assigns a goal rather than a measured result.
On Saturday Altman told Fortune that OpenAI would forgo a 2026 listing on safety grounds, and he endorsed a proposal from Anthropic chief executive Dario Amodei for the industry to slow gains in model capability.
OpenAI has made no public comment on Bel.
