According to a leak, OpenAI has completed the pretraining of a new base model codenamed Bel. The model has more than ten trillion parameters, is roughly the same size as GPT-4.5, and will succeed the Doug model. It will serve as the basis for Astra and, after further reinforcement learning, GPT-6 as well. The information appeared on X.
What Bel Is Supposed to Be
According to the post, OpenAI does not plan to deploy the Bel model immediately. The company sees it as a foundation for systems coming after GPT-6 and acknowledges that it could evolve into a model on the threshold of artificial general intelligence. Wccftech, which analyzed the report in greater detail, described it as wild speculation circulating on social media.
The difference between pretraining and a finished product is significant. At this point, Bel is a raw foundation—an enormous network trained on text that still requires further fine-tuning. Only then will it become a product that users can access through an application. Some commentators estimate that Bel could appear before the end of this year, but this has not been confirmed, and development may take longer.
🚨 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 Road to Bel Led Through Garlic and Doug
To understand the name Bel, it is necessary to go back two years. OpenAI released GPT-4o in May 2024. Since then, although it launched GPT-4.5 among other models, it has not completed a full pretraining cycle that could serve as a new generation of its flagship model. Instead, improvements in capabilities were driven by reinforcement learning and the computing power used during the model's actual operation, beginning with o1-preview in September 2024. The analytics firm SemiAnalysis claims that advances in both o3 and the entire GPT-5 series were still built on foundations dating back to the GPT-4o era.
However, this approach is reaching its limit. When Google released Gemini 3 on November 18, 2025, the issue of technical direction turned into direct market pressure. On December 1, the media reported that OpenAI CEO Sam Altman had declared an internal state of emergency called Code Red and ordered resources to be shifted toward improving ChatGPT. The very next day, The Information reported that the company was training a new base model called Garlic. This model excelled at programming and reasoning and benefited from fixes to errors discovered during earlier attempts. OpenAI's head of research, Mark Chen, told the team that they had managed to resolve some of the old pretraining issues. As a result, smaller models were able to absorb knowledge that had previously required far more parameters.
According to the same report, the company was already building a larger model based on insights gained from Garlic. That successor was Doug. According to a report by SemiAnalysis, OpenAI had already overcome its pretraining difficulties by then and was working intensively on a much larger model codenamed Doug. Two days later, this was confirmed by leaker ChrisGPT, who described it as the largest pretraining project in the company's history. The Doug model is reportedly not the same as GPT-6 and could launch by November of this year at the latest. Bel therefore follows Doug, which is based on Garlic. It is the third iteration of the same effort, but each time on a different scale.
Meanwhile, the Wait for Astra Continues
Meanwhile, the public has not yet gained access to the model at the center of it all. OpenAI revealed the name Astra in an unconventional way. It published a research paper containing ten solutions to long-standing open problems in mathematics and theoretical computer science, accompanied by manuscripts and certificates written in the Lean language. The company said that an internal version of the model had solved problems that specialists had been unable to resolve for at least ten years.
According to the description, Astra moves beyond the one-question, one-answer model. A supervising agent breaks a complex problem down into smaller parts and distributes them among internal subagents. It then combines their findings. It may work on a single task for hours or even days.
However, no release date has been announced. According to published reports, Astra was supposed to undergo a new federal review before its release, a process that takes roughly a month. OpenAI subsequently announced, however, that it was suspending activities that did not meet stricter security checks. On August 7, the company admitted that, under its own safety framework, it could not rule out the upcoming model reaching the critical threshold for cyber capabilities. It was the first such case in the company's history.
Altman nevertheless insists that the company will make Astra available to the public. He considers reserving powerful models for a select few to be the wrong path.
Bad News for Anthropic
The leak was not limited to parameter counts. According to the source, OpenAI believes that Anthropic has no adequate public response to Astra ready. The author of the post adds that his information confirms this. The main reason is a lack of computing capacity. Anthropic reportedly expects the rest of the year to favor OpenAI and hopes to return to the top early next year.
Anthropic pays SpaceX roughly fifteen billion dollars per year for computing power. OpenAI, meanwhile, operates at a gross margin of around seventy percent and has virtually unlimited capital for compute. According to Polymarket, Anthropic had a seventy-four percent chance of releasing a new flagship model within two weeks, most likely Mythos or Fable 5.1.
However, not everything is going well for OpenAI either. On the day information about the Doug model leaked, its ChatGPT Atlas browser was discontinued after less than ten months of operation. Fidji Simo, head of the applications division, had already stated clearly in March that the company was scaling back side projects and shifting resources toward programming and enterprise deployments. The discontinuation of both the Sora video generator and the Atlas browser is a continuation of this wave of streamlining.
Source: finance.biggo.com
