OpenAI Reportedly Completes Bel Pretraining With More Than 10 Trillion Parameters
OpenAI has reportedly completed pretraining for an enormous new frontier base model internally codenamed Bel, with claims suggesting it contains more than 10 trillion total parameters and could underpin models beyond the upcoming Astra and GPT 6 generation. The information originates from AI industry source @synthwavedd, who describes Bel as the successor to an earlier internal pretraining project known as Doug. OpenAI has not publicly confirmed Bel, its parameter count, architecture, training compute or intended product roadmap, making every technical detail surrounding the model unverified for now.
According to the leak, Bel exceeds 10 trillion total parameters and is roughly comparable in overall parameter scale to GPT 4.5. The model is reportedly intended to provide a foundational pretraining layer from which OpenAI can extract multiple increasingly capable systems through reinforcement learning and additional post training. Astra and GPT 6 class systems are specifically associated with that pipeline in the report, while Bel could potentially continue serving as a foundation for models developed after GPT 6.
🚨 SCOOP: OpenAI recently finished its next pretrain, codename "Bel" - the successor to "Doug", which is expected to be the base for Astra and GPT-6 (w/ further RL)
— leo 🐾 (@synthwavedd) August 25, 2026
Bel is a giant pretrain with >10T total parameters, similar in size to GPT-4.5. OpenAI expect it to be a…
The most aggressive claim is that OpenAI internally views Bel as potentially capable of supporting what the source describes as an "AGI threshold" model. That wording should be treated cautiously. There is no universally accepted technical benchmark that determines when a model becomes AGI, and parameter count by itself does not establish general intelligence. Architecture, training data, inference compute, reinforcement learning, tool use, memory, agent design and post training can all dramatically influence capability. OpenAI itself currently describes AGI broadly as systems capable of solving human level problems, but it has made no public statement connecting Bel with that threshold.
What is confirmed is that OpenAI continues scaling both its models and the infrastructure behind them. The company recently released GPT 5.6 and has publicly discussed increasingly capable unreleased systems during its scientific research, including models that have produced substantial progress on long standing mathematics and theoretical computer science problems. OpenAI is simultaneously expanding Stargate and has described the project as part of the compute foundation required to pursue increasingly advanced intelligence. None of those official disclosures, however, confirm that the underlying system is Bel.
The reported training milestone also arrives as OpenAI becomes increasingly vertically integrated around AI infrastructure. The company has been expanding its own silicon strategy alongside partners, an effort previously explored through our coverage of OpenAI's first custom Jalapeño inference chip. If Bel really operates at the reported scale, the compute required not only to train it but eventually serve descendants of it would reinforce why OpenAI is investing simultaneously in data centers, custom accelerators and large external compute partnerships.
A 10 trillion plus parameter pretraining run would be a massive technical milestone, but the parameter number should not become the story by itself. Modern frontier AI increasingly depends on what happens after pretraining, where reinforcement learning, agentic systems, tool integration and inference time compute can extract far more useful capability from the underlying model.
The "AGI threshold" description is therefore the most interesting part of the rumor and also the least verifiable. Until OpenAI publishes benchmarks, architecture information or an official model announcement, Bel should be viewed as a potentially important internal foundation rather than evidence that AGI has arrived. If the leak proves accurate, however, OpenAI may have completed the base model that powers several generations of its next frontier systems.
Would a 10 trillion parameter model represent a meaningful step toward AGI, or have architecture and post training become more important than raw parameter count?
