Is OpenAI Already Training What Comes After GPT-6? Inside the "Bel" Leak Nobody Can Verify

OpenAI shipped GPT-6 "Astra" to the public barely six weeks ago, in early September. According to a rumor that has been circulating on X since late August, the company had already finished training its next foundation model before Astra even reached users — a long shot larger system reportedly codenamed "Bel." Treat this one carefully: it traces back to a single anonymous account, OpenAI has not acknowledged it in any form, and the outlets repeating it disagree on basic details.

Where the rumor actually comes from

The claim originates from a post by the X account @synthwavedd on August 25, 2026, which asserted that OpenAI had completed a pretraining run codenamed Bel with more than 10 trillion parameters — reportedly the successor to an earlier, equally unconfirmed run called "Doug." The post was then picked up and amplified by outlets including Wccftech, WinCentral and China's 36Kr, several of which ran headlines flatly stating Bel "beats" GPT-6 Astra. That specific comparison, worth noting, doesn't appear to come from the original post at all — it looks like an embellishment added somewhere in the retelling.

As of this week, "Bel" doesn't appear anywhere in OpenAI's model documentation, its API changelog, or any statement from an OpenAI employee. That absence doesn't disprove the rumor — frontier labs routinely train models internally for months before anyone outside the building hears a name — but it does mean there is currently no independent confirmation beyond one social post and the outlets that reran it.

The one piece of the story with slightly more backing

The earlier codename in the chain, "Doug," has a bit more going for it. SemiAnalysis, a research outlet with a track record of sourcing from people close to frontier AI labs, reportedly referenced a much larger in-progress OpenAI model under that name in a subscriber note. That's a firmer data point than an anonymous post, but it's still secondhand, still unconfirmed by OpenAI, and says nothing definitive about Bel specifically or about a 10-trillion-parameter figure.

It's also worth flagging that even if the parameter count holds up, it may not mean what it sounds like. Modern frontier models increasingly use sparse mixture-of-experts architectures, where only a fraction of parameters activate for any given query. A "10 trillion parameter" headline number, if accurate, would not necessarily translate into 10 trillion parameters' worth of active compute per response — and parameter count alone has never reliably predicted real-world capability.

What's actually confirmed, for comparison

To be clear about what we do know: GPT-6 Astra itself is real and shipped on schedule, reaching approved enterprise organizations on September 3 and rolling out to paying ChatGPT subscribers the following day. OpenAI also reportedly restarted a large reinforcement-learning training run in late August, and by September 8 the company was said to be crediting an internal model that outperforms the public GPT-6 Astra on certain internal benchmarks — though, again, "internal model" is a long way from a confirmed Bel, Doug, or GPT-7.

In other words: it's well established that OpenAI, like every major lab right now, has multiple models in flight at different stages behind the one you can actually use. Whether any of them is specifically called Bel, whether it's finished, and whether it's 10 trillion parameters is the part still resting entirely on one unverified post.

Why this matters in the Gulf specifically

Rumors about OpenAI's internal roadmap aren't purely a Silicon Valley story for the UAE. Abu Dhabi is currently host to Stargate UAE, the 1-gigawatt AI data center cluster that G42 is building as part of a larger 5-gigawatt UAE-US AI campus, with OpenAI and Oracle operating the compute and Nvidia, Cisco and SoftBank also involved. The project was targeting roughly 200 megawatts of capacity online sometime in the third quarter of 2026; as of the most recent public tracking, it's not confirmed whether that first phase has actually gone live yet.

That matters here because training (and eventually serving) a model of the scale rumored for Bel — if it exists — requires exactly the kind of gigawatt-scale compute Stargate UAE was built to provide. Whatever comes after GPT-6, some meaningful share of the infrastructure behind it is reportedly being built in Abu Dhabi, not just in the US. That's a real, separately-confirmed fact, distinct from anything in the Bel rumor itself.

The bottom line

Nobody outside OpenAI has confirmed that a model called Bel exists, that it finished training, or that it has 10 trillion parameters. What's confirmed is that GPT-6 Astra is out, that OpenAI is very likely training something bigger behind the scenes (every frontier lab is, at any given moment), and that a serious chunk of the compute that could eventually run it is being poured into the ground in Abu Dhabi right now. Until OpenAI says otherwise, file "Bel" under plausible-but-unverified, and watch whether SemiAnalysis or another sourced outlet corroborates the name independently in the weeks ahead.