We are now getting leaks about GPT-7 for a product line where GPT-6 does not exist.

On August 25, an X account posted what it called a scoop: OpenAI had finished a new pretraining run codenamed Bel, successor to an earlier run called Doug, with more than 10 trillion total parameters. Bel was described as the foundation for Astra and GPT-6 with further reinforcement learning, and as a potential base for what comes after, possibly an AGI-threshold system.

Within days it was everywhere. Wall St Engine picked it up. Tech outlets ran it. And one widely shared post asked the obvious question out loud: what will OpenAI call the model built on Bel? GPT-7? Already?

Here is what the leak actually says, where it contradicts itself, and the one detail buried in it that is more interesting than the parameter count.

The core claims are consistent across the versions circulating.

OpenAI has reportedly completed a pretraining run codenamed Bel, exceeding 10 trillion total parameters, which the leak describes as similar in size to GPT-4.5. It is said to follow an earlier run called Doug. And it is positioned as a base model rather than a product, meaning it would need substantial post-training before anything ships.

The ambitious part is the framing. The leak says OpenAI expects Bel to serve as a post-GPT-6 base, and that with further reinforcement learning it could underpin a system at what people are calling the AGI threshold.

One correction worth making immediately, since the headlines blur it: Bel is not GPT-7. Even the leak does not claim that. It describes a raw pretraining foundation, the thing you build a product on top of, not the product. The GPT-7 question is what the community added on top.

This is the part most coverage skipped, and it matters more than any of the specs.

The lineage is inconsistent. Some versions say Bel is Doug's successor. Others describe Bel as the foundation for the Astra and GPT-6 generation, which would put it earlier, not later. Those are different claims and both are circulating from anonymous accounts. One tracker flagged the contradiction directly, noting the lineage across these accounts does not line up.

The number is already inflating. The original claim was more than 10 trillion parameters. A separate wave of social amplification pushed that to 100 trillion. Nobody added evidence in between. That is a leak growing by an order of magnitude purely through retelling, in about a week.

It may never be verifiable. OpenAI has historically confirmed base models through launch events with published system cards, and even then it rarely discloses parameter counts. So a 10-trillion figure may never be directly checkable against anything official, which means it can circulate indefinitely without ever being resolved.

And nothing is confirmed. No OpenAI announcement, no model card, no technical paper, no benchmark, no pricing, no access method. Everything traces back to posts on X and their pickup by the tech press.

Strip out the AGI talk and there is a genuinely interesting claim in the Doug half of this.

Reporting describes Doug as OpenAI's return to scaling base models after roughly sixteen months without a genuinely new foundation model, a stretch in which the company leaned mostly on reinforcement learning and inference-time compute, extra thinking spent while the model answers rather than while it trains.

If that is accurate, it says something concrete about strategy. The industry spent this year arguing that scaling had stalled and that gains would come from post-training instead, an argument DeepSeek and Zhipu both supported by posting large jumps from re-post-training unchanged base models. Doug and Bel would represent OpenAI betting the opposite way, that pure pretraining scale still has room, and spending enormous compute to prove it.

One leaker put Doug's estimated launch around November 2026, which is the only date attached to any of this and comes with no more evidence than the rest.

Buried in the original thread is a competitive claim that got less attention than the parameter count and is far more falsifiable.

The leaker wrote that OpenAI believes Anthropic has no good response to Astra prepared for a public launch, largely because of compute constraints, and that Anthropic is bracing for much of the rest of the year to go OpenAI's way while expecting to be back on top early next year.

Treat that as what it is, an anonymous account characterizing one company's belief about another. But unlike a parameter count, it is testable on a timeline. Anthropic is targeting an IPO in the coming weeks and has gone unusually quiet on model releases. Whether it ships something significant before year end is a claim that resolves itself, and soon.

The stories moving fast beyond today's headlines:

So is Bel GPT-7? Maybe. Nobody knows, including possibly OpenAI.

Here is the honest position. A base model this large has to become something. If Astra ships as GPT-6, then whatever gets built on Bel is the next generation, and GPT-7 is a reasonable guess at the name. That is speculation, but it is not stupid speculation.

What makes it shaky is everything around it. GPT-6 is not a confirmed name. Astra is paused over cyber concerns with the largest frontier run still on hold. And the parameter count went from 10 trillion to 100 trillion in a week through pure retelling.

Nothing here changes what you build. There is no model, no name, no date. Anyone selling GPT-7 access is selling nothing.

Three things worth watching instead of the rumors: whether Doug, Bel, or Astra ever appear in official OpenAI material, how fast new data center capacity comes online, and what OpenAI actually names and prices when it ships next.

Stay building. 🤖