At 2:00 AM in Jakarta on 26 August, one X post named an alleged OpenAI pretrain called Bel. Five hours later, the claim had become a roadmap, a giant star, and a conclusion about Anthropic. In those five hours, one unsupported claim became an entire model roadmap.
The post before the viral screenshot
The first explicit public version found came from Leo, @synthwavedd. His X post called itself a “SCOOP” and made a long chain of claims in one pass: OpenAI had finished a new pretrain named Bel; Bel followed an internal project called Doug; Doug would become the base for Astra and GPT-6 after further reinforcement learning; and Bel contained more than 10 trillion total parameters.
The same post went further. It compared Bel's size with GPT-4.5, described it as a possible base for a model near an “AGI threshold,” and attributed a private competitive view to both OpenAI and Anthropic. It linked no document, named no source, quoted no employee, and attached no technical record. The next post in the thread promoted a Discord server for earlier access to leaks.
The screenshot that later circulated more widely came from Adit_Yah, @Adidotdev, at 7:08 AM Jakarta time. Adit's version changed “recently finished its next pretrain” into “already finished the model after Astra,” added the purple-star-versus-orange-star illustration, and concluded that Stargate was paying off for a new training wave. Five seconds later, Adit posted cc: @synthwavedd with a screenshot of Leo's post. The credit resolves the source chain: Adit was a restatement, not a second leak.
Repetition did not become corroboration
The claim travelled quickly. A Reddit post appeared nine minutes after Leo's and began with “According to Leo”. Wccftech later ran a headline stating that Bel had more than 10 trillion parameters and might be an AGI-threshold base model. Its article, however, embedded the same X post and called the story a “wild rumor.”
The distribution trail is broad; the evidence trail ends at Leo. Reddit, Adit's post, the Wccftech story, and later copies all point back to him. None provides an authenticated OpenAI artifact or independently sourced reporting. The safest description is therefore precise: Leo is the earliest traceable public source found, and Bel remains his unconfirmed claim.
Private Discord messages, deleted posts, and unindexed communities remain outside the searchable record. That boundary prevents an absolute claim about who first said “Bel” anywhere. It does not rescue the public story from its single-source problem.
What OpenAI actually confirms about Astra
OpenAI has confirmed Astra. On 1 August, it said an internal version of Astra, “our next major model,” produced results on ten mathematics and theoretical computer-science problems. The official publication describes the problems, manuscripts, and Lean certificates. It does not mention Bel, Doug, GPT-6, a parameter count, or a launch date.
Six days later, OpenAI called Astra “one of our upcoming models” while explaining new controls around its possible cyber capability. On 18 August, the company said it had temporarily slowed scaling, paused RL training on its latest deployment-intended models for two weeks, and kept its largest planned frontier RL run on hold while smaller work continued. It also said a significant number of Astra workloads remained paused under the new security bar. The status update supports “upcoming.” It does not support a confident “near-term launch.”
This does not disprove a completed secret pretrain. Pretraining, post-training, evaluation, safety work, and deployment can occupy different stages. It does mean the Doug→Astra→GPT-6→Bel sequence cannot be read out of OpenAI's public record. The earlier Astra origin investigation drew the same basic line: Astra was public; the GPT-6 label was not. The Bel post adds a new rumor on top of that line, not a correction to it.
“10T” changed meaning as the post travelled
Leo's exact wording was “>10T total parameters.” It did not say training tokens. Adit shortened that to “10T+ pretrain,” which is easier to repeat and easier to misread. Parameters and training tokens are different units. Total parameters, active parameters per token, data tokens, and training compute are different measurements again.
The GPT-4.5 comparison cannot repair the gap. OpenAI's GPT-4.5 launch page called it the company's largest GPT chat model at release and later described it as very large and compute intensive. OpenAI did not publish a parameter count there. Without a public number for GPT-4.5 and a verified number or architecture for Bel, “same size class” is not a comparison anyone can reproduce.
Nor does a completed pretrain, if one exists, imply a finished product. The same GPT-4.5 page separates scaled pretraining from supervised fine-tuning and RLHF. OpenAI's August update shows how later training and evaluation can be paused or resumed independently. The viral star illustration collapses all of those unknowns into two circles. It is effective rhetoric and unusable measurement.
Stargate is real training infrastructure—and still not proof
Stargate gives the rumor a real piece of infrastructure to point at. In January 2025, OpenAI and SoftBank announced an intent to invest $500 billion over four years, with $100 billion to begin deploying immediately. Those were commitments, not a statement that the full capacity was already online.
By July 2025, the record included live work. OpenAI and Oracle announced an additional 4.5 gigawatts, bringing Stargate to more than 5 gigawatts under development. In the same update, OpenAI said parts of its Abilene campus were operating and early training and inference workloads had begun. Reuters independently confirmed the agreement while noting that locations and funding details for the new facilities were not then disclosed and that analysts had questioned the venture's funding.
OpenAI's September expansion put Stargate at nearly 7 gigawatts of planned capacity, while again describing Abilene as live and running early workloads. “Planned,” “under development,” and “operational” are three different states.
Stargate had already supported model training. No checked source identifies those workloads as Bel, Doug, a 10-trillion-parameter system, or the base of GPT-6. Adit added the Stargate conclusion; Leo's source post did not contain it. Available compute can make a large run plausible. It cannot authenticate a codename.
The Anthropic passage asks for evidence it never supplies
Leo's post says OpenAI believes Anthropic has no good public answer to Astra, mostly because of compute constraints, and that Anthropic expects much of 2026 to favor OpenAI before a comeback in early 2027. These are claims about private beliefs, private capacity judgments, and a private roadmap. They carry a higher burden than a product rumor. The post names no OpenAI or Anthropic speaker and points to no document or interview.
The public record does not establish who would win an Astra comparison, because Astra has no public access or matched benchmark. It does show that Anthropic was not waiting quietly for 2027. Anthropic launched Opus 5 on 24 July. Reuters confirmed the release and quoted product leader Dianne Penn saying it reflected a rapid pace of development.
Anthropic also announced an Amazon agreement for up to 5 gigawatts of compute, said it already used more than one million Trainium2 chips, and expected nearly 1 gigawatt of Trainium2 and Trainium3 capacity online by the end of 2026. Those are company disclosures, not an audited head-to-head comparison with Stargate. They are enough to show why “mostly compute” needs data rather than assertion.
Opus 5 is not a proven answer to Astra. The claim that Anthropic had accepted a weak 2026 and planned a 2027 return has no named evidence behind it.
The verdict, claim by claim
| Claim | Verdict | Reason |
|---|---|---|
| Astra is an upcoming OpenAI model | Verified | OpenAI calls it its next major and upcoming model. |
| OpenAI completed a pretrain called Bel | Claimed | One X post says so; no OpenAI artifact or independent confirmation was found. |
| Doug becomes Astra or GPT-6 after RL | Unknown | The relationship appears only in the same post. |
| Bel has more than 10T parameters | Claimed | The number has no attached technical record. |
| Bel matches GPT-4.5's size class | Not checkable | OpenAI did not publish GPT-4.5's parameter count. |
| Stargate runs training workloads | Verified | OpenAI says early workloads ran at Abilene. |
| Stargate trained Bel | Unsupported | No checked source joins the infrastructure to that codename. |
| Anthropic accepted 2026 and planned a 2027 comeback | Unsupported | No named speaker, document, or direct report supports it. |
What would move Bel out of the rumor column
An authenticated OpenAI model card, paper, repository record, executive statement, or training artifact naming Bel would change the verdict. So would independent reporting that names its sourcing, describes direct access, and ties the codename to a specific run. A correction from Leo that exposes the original evidence could also change the chain.
Another screenshot, a larger view count, or an article that embeds the same post would not. Neither would another Stargate expansion without a model-level link.
For the next Bel update, the shortest audit starts with Leo's original post, then asks whether the new evidence ever leaves that source chain. Until it does, the clean verdict remains: Astra is confirmed, Stargate is training models, and Bel is a one-source claim.
Loading reader comments...