OpenAI may already be working several steps beyond GPT-6, if a new leak about an internal project codenamed “Bel” turns out to be accurate.
According to unverified reports circulating online, Bel is believed to be a major OpenAI pretraining run that has already been completed. The project reportedly follows an earlier internal effort known as “Doug” and could form the foundation for the company’s next generation of models.
The claims have quickly attracted attention because of one particularly ambitious detail: Bel is rumored to involve more than 10 trillion total parameters.
Bel Is Reportedly Not GPT-7
Despite the GPT-6 comparisons, Bel is reportedly not being positioned as GPT-7 itself.
Instead, the leaked information suggests that Bel could serve as a foundational model or pretraining layer for whatever comes after GPT-6. In other words, OpenAI could potentially be training the underlying technology for its next major generation before GPT-6 has even fully arrived.
That would fit with how frontier AI development increasingly works. Companies don’t necessarily wait for one model generation to launch before beginning work on the technologies that will power the next one.
The AGI Angle
The most eye-catching part of the rumor is the claim that some insiders allegedly view Bel as OpenAI’s first serious attempt at an AGI-threshold model.
That’s a much bigger claim than simply suggesting a larger language model is being trained. It implies that Bel could be designed around capabilities that go beyond today’s increasingly powerful chatbots and coding assistants.
However, there is currently no public confirmation from OpenAI that Bel exists, that training has been completed, or that the project has anything to do with AGI.
The reported parameter count should also be treated cautiously. Parameter counts are becoming a less useful way of comparing modern AI systems, particularly as companies increasingly rely on techniques such as mixture-of-experts architectures and sophisticated post-training.
Astra Isn’t Even Fully Rolled Out Yet
The timing makes the rumor even more interesting.
OpenAI’s Astra project is reportedly focused on creating a much more capable AI assistant that can interact with computers, work across applications, analyze information and carry out longer-running tasks.
Yet if the Bel reports are accurate, OpenAI may already have the foundation for an even more ambitious generation of models behind the scenes.
That would suggest Astra could be only one step in a much larger roadmap rather than the endpoint of OpenAI’s current AI push.
What We Know—and What We Don’t
For now, Bel remains an unverified leak. There is no official OpenAI announcement confirming the codename, the reported 10T+ parameter figure, its relationship with Doug, or claims that it represents an AGI-threshold model.
Still, the idea is notable because it highlights how far ahead frontier AI companies may be working internally.
If the reports eventually prove accurate, the biggest story may not be that OpenAI has already trained a model after GPT-6. It could be that the company is developing the foundation for its post-GPT-6 era while the next generation hasn’t even fully arrived.
Note: The claims about Bel, Doug, the reported parameter count and AGI positioning are based on unverified reports and should not be treated as confirmed OpenAI information.
For continuing updates on OpenAI, ChatGPT, Codex, GPT models, and AI industry developments, bookmark:
