OpenAI has reportedly completed a massive pretraining run codenamed "Bel," a model said to feature more than 10 trillion parameters. If the claim proves accurate, Bel would represent the company's first successful full-scale frontier pretraining since GPT-4o shipped in May 2024. For context on that scale: OpenAI's last officially disclosed parameter count dates back to GPT-3, listed at 175 billion parameters in its 2020 paper, and the company has not published figures for any model since. Frontier labs that have disclosed very large totals typically rely on sparse mixture-of-experts designs, in which only a fraction of a model's parameters activates for any given token — DeepSeek-V3, one of the largest publicly documented models, reports 671 billion total parameters with roughly 37 billion active per query.

The report originated from a post on X by user @synthwavedd on August 25, 2026. OpenAI has not officially confirmed that Bel exists, and the company has released no technical specifications for the run.

What is known about Bel

According to the unverified reports, Bel follows a prior pretraining run codenamed "Doug" and is intended to serve as a base model for future systems. Those future systems are said to include a project known as Astra, as well as successors to GPT-6, with post-training work expected to unlock capabilities that push closer to what the industry calls Artificial General Intelligence. Pretraining is the phase in which a model absorbs knowledge from vast text corpora before post-training techniques shape its behavior, which is why the base model is generally treated as the foundation that bounds what later systems can extract from it.

OpenAI has reportedly struggled with scaling since GPT-4o launched. Multiple sources within the AI research community have pointed to interrupted training runs, safety-related pauses in reinforcement learning processes, and technical difficulties that stalled progress on larger models. The company's compute scaling efforts have been discussed under the umbrella of Project Stargate — the initiative announced in January 2025 with SoftBank, Oracle, and MGX as partners, which plans up to $500 billion in U.S. data-center investment over four years beginning with a campus in Abilene, Texas. That infrastructure ambition reflects how compute-hungry frontier training has become: published estimates, including figures compiled for the Stanford AI Index, have placed individual frontier training runs at tens of millions of dollars to nearly $200 million.

Why the gap since GPT-4o matters

The period between May 2024 and now has been an unusually long stretch for OpenAI to go without shipping a new frontier base model. The company released incremental updates and fine-tuned variants during that time — including GPT-5 in August 2025 and the point releases that followed — but its core pretraining work had not produced a confirmed successor at a meaningfully larger scale. Under the report's framing, those shipped models drew on base work trained earlier rather than on a new, larger pretraining run.

Competitors did not stand still during the same window. Anthropic pushed forward with its Claude model family, Google DeepMind continued scaling Gemini, and a growing roster of open-source efforts from Meta and others kept closing the gap.

The AGI question and competitive stakes

The framing around Bel is explicitly tied to AGI ambitions. The reported pipeline — from Bel to Astra to whatever follows GPT-6 — suggests OpenAI is thinking in terms of a multi-stage roadmap in which each layer of post-training and fine-tuning extracts progressively more capable behavior from the base model.

The cautionary note is equally straightforward: none of this has been confirmed. Until OpenAI provides official validation — whether through a technical paper, a product announcement, or even a casual blog post — Bel remains an unverified claim originating from a social media post. The practical checkpoints for readers are concrete. OpenAI has historically confirmed major base models through launch events accompanied by published system cards, and even then it rarely discloses parameter counts, so a 10-trillion figure may never be directly verifiable from official sources. Other observable signals include the pace at which Stargate capacity comes online, any appearance of the reported codenames — Doug, Bel, Astra — in official materials, and the timing and positioning of whatever OpenAI ships next.