OpenAI has reportedly completed a next-generation pretraining run codenamed Bel, according to an August 25, 2026, post by @synthwavedd. The post describes Bel as the successor to Doug, with more than 10 trillion total parameters, and says it is expected to serve as the base for Astra and GPT-6 after additional reinforcement learning (RL).

Original post on X

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

Bel is positioned as a foundation for future models

The reported roadmap places Bel upstream of multiple OpenAI systems rather than presenting it as a finished public product. According to the post, Astra and GPT-6 would be built from the completed pretraining run and then receive further RL before release.

The post also says Bel could become the base for a model released after GPT-6, potentially including a system the author associates with an AGI threshold. That is presented as a possibility and an expectation about OpenAI’s plans, not as a confirmed launch announcement or a claim that Bel itself has reached that threshold.

The reported scale of the pretraining run

The post characterizes Bel as a large pretraining effort with more than 10 trillion total parameters and compares its size with GPT-4.5. Its significance, if the report is accurate, would come not only from the claimed scale but also from the number of future systems it is expected to support.

Because Bel is described as a base model requiring additional RL, completion of the pretraining stage does not by itself indicate that Astra or GPT-6 is ready for public access. The post is reporting a development milestone in a model pipeline, not announcing a product launch.

The competitive forecast involving Anthropic

The post connects the reported milestone to competition between OpenAI and Anthropic. It says Anthropic is concerned about Astra and is not believed to have a strong response prepared for a public launch. It attributes that position mainly to compute constraints and forecasts that the rest of 2026 could favor OpenAI.

It also says Anthropic expects to regain the lead early in 2027. These are competitive assessments and a forward-looking forecast contained in the post, rather than confirmed product roadmaps announced by either company.

The central claim is therefore about OpenAI’s reported training progress and the possible role of Bel as a foundation for several future models. The details about Astra, GPT-6, a post-GPT-6 system, and Anthropic’s expected position should be read as reported expectations until the companies provide their own announcements.