OpenAI has not announced a model called Bel, and nothing it has published uses the name. What exists is one X post from August 25, 2026, a September 5 relay that turned it into “GPT-6 Bel Leaks: Beats GPT-6 Astra”, and a month of repetition. This page grades each claim against what OpenAI has actually published, as of October 4, 2026.
We checked OpenAI’s model docs, API changelog and news feed at 00:40 UTC on October 4: zero mentions of Bel and zero of Doug.
On this page · 10 sectionsOpen
- OpenAI has not confirmed Bel. As of October 4, 2026 the name appears nowhere in OpenAI’s model docs, API changelog or news feed, and no OpenAI employee has used it in public.
- Bel comes from one X account, @synthwavedd, on August 25, 2026: a finished pretrain with more than 10 trillion total parameters, the successor to an earlier run called Doug. That post has drawn about 811,000 views.
- Doug has an older paper trail. On August 7 SemiAnalysis published an excerpt from a July 9 subscriber note saying a much larger OpenAI model codenamed Doug was in the works.
- The viral ‘GPT-6 Bel beats GPT-6 Astra’ headline adds a claim the source never made: no benchmark exists, and the original post called Bel a base for models after GPT-6.
- OpenAI’s own dated facts are different: it restarted a large reinforcement learning run on August 28, and on September 8 credited an internal model more capable than GPT-6 Astra. It has called neither Bel.
- The same account wrote on September 23 that Bel is still months from going public. On October 3, OpenAI’s Codex lead teased a release next week without naming a model.
§ 01What we actually know
| Claim | Grade | Why |
|---|---|---|
| OpenAI finished a pretrain codenamed Bel | Unconfirmed | One X account; no OpenAI document or post |
| More than 10 trillion total parameters | Unconfirmed | Same post; OpenAI has published no figure |
| Bel succeeds an earlier run called Doug | Partly corroborated | Doug appears independently in a SemiAnalysis note dated July 9 |
| Doug is the base for GPT-6 Astra | Unconfirmed | OpenAI has not described Astra’s base model |
| Bel is meant as a post-GPT-6 base | Unconfirmed | The poster’s account of OpenAI’s expectations |
| Bel, not Astra, would be the GPT-6 jump | Contradicted | OpenAI released GPT-6 Astra on September 3, a day after this reading |
| GPT-6 Bel beats GPT-6 Astra | No source | Added by a relay headline on September 5; no benchmark exists |
| OpenAI trained a model more capable than Astra after August 28 | Confirmed, not named Bel | OpenAI’s September 8 Navier-Stokes post |
| OpenAI releases something next week | Teased, unnamed | A reply by Codex lead Thibault Sottiaux on October 3 |
§ 02The posts, word for word
The source is @synthwavedd, at 19:00 UTC on August 25: “OpenAI recently finished its next pretrain”, followed by the codenames. The size claim reads: “Bel is a giant pretrain with >10T total parameters, similar in size to GPT-4.5.” And the expectation: “OpenAI expect it to be a post-GPT-6 base, and potentially even the base for an AGI-threshold model.” The post cites no document, and GPT-4.5’s own size was never published by OpenAI, so the comparison is a second estimate.
On September 2 the same account, citing The Information on Astra’s naming, wrote that “it suggests they’re planning for Bel to be the 6-worthy jump” (post). OpenAI shipped GPT-6 Astra the next day, so that reading did not hold.
On September 23 it set the clock: Astra’s compressed timeline helps Bel, “but it’s still ~months out” (post).
The viral version came from @Mr_Salio on September 5, about 298,000 views. Its headline, “GPT-6 Bel Leaks: Beats GPT-6 Astra”, carries a comparison its own bullet points never support: the bullets repeat the August 25 claims and nothing else.
§ 03The chain, dated
| Date (UTC) | Who | What | Views on X |
|---|---|---|---|
| July 9 | SemiAnalysis, subscriber note | A much larger OpenAI model codenamed Doug is in the works | Not public until Aug 7 |
| August 7 | SemiAnalysis newsletter | Publishes the July 9 excerpt | Newsletter |
| August 25, 19:00 | @synthwavedd | Bel finished, over 10T parameters, successor to Doug | 811,976 |
| August 28 | OpenAI | Restarts a large frontier RL run | OpenAI post, Sep 1 |
| September 2 | @synthwavedd | Bel as the 6-worthy jump | 595,654 |
| September 3 | OpenAI | Ships GPT-6 Astra | Launch |
| September 5 | @Mr_Salio | GPT-6 Bel beats GPT-6 Astra headline | 298,599 |
| September 23 | @synthwavedd | Bel still months out | 37,184 |
| September 28 | WSJ, OpenAI | GPT-6.1 Astra scrapped | News |
| October 3, 05:38 | Thibault Sottiaux | A release next week, unnamed | 453,208 |
§ 04Where Doug comes from
Doug is the better-sourced half. SemiAnalysis’s August 7 newsletter quotes a note it sent Tokenomics Model subscribers on July 9: “OpenAI has overcome their pre-training issues” and a much larger model codenamed Doug “is actively in the works”. That is an analyst firm’s claim, not OpenAI’s, but it predates and stands apart from the Bel post. The Bel post’s line that Doug is “expected to be the base for Astra and GPT-6” has no second source.
§ 05What OpenAI has said, and what it was about
OpenAI’s dated training facts concern reinforcement learning, not pretraining. Its September 1 Path to Astra post says: “On August 28th, we restarted the large frontier RL run that was previously paused”. A pretrain is the base model; an RL run trains a model further. Treating August 28 as Bel’s date mixes the two.
On September 8 OpenAI’s Navier-Stokes post credited an internal model it called significantly more capable than GPT-6 Astra, in training since August 28. @synthwavedd described it the same day as “an internal model seemingly a generation ahead of Astra”. OpenAI did not name it, and did not say what base it uses.
Then, on September 28, OpenAI scrapped GPT-6.1 Astra after alignment tests; per the WSJ it plans more reinforcement learning on the same base model. Nothing there mentions a new pretrain.
§ 06The next-week tease
At 05:38 UTC on October 3, Thibault Sottiaux, who leads Codex, replied to a user: “You don’t know what we’re releasing next week”. It has drawn about 453,000 views and names nothing. Read against the September 23 “months out” post from Bel’s own source, a Codex release next week is far more likely to be a Codex or Astra-line update than a new base model. That is a reading, not a fact; this page changes if OpenAI’s post says otherwise.
§ 07What the markets say
| Market | Outcome | Probability |
|---|---|---|
| Manifold, GPT 7 (OpenAI) release date | Before January 2027 | 3.9% |
| Polymarket, GPT-7 released by | June 30, 2027 | 36.5% |
| Polymarket, GPT-7 released by | December 31, 2027 | 88.5% |
Markets measure sentiment, not evidence, and none asks about Bel by name. They agree with the source on one thing: whatever Bel becomes is not expected this year.
§ 08What would confirm it
- An OpenAI post, system card or model page naming Bel or a successor base model.
- A parameter count or training-compute figure from OpenAI for any model after GPT-6 Astra.
- OpenAI’s next-week release, read for what base it says it uses.
- A second outlet with its own sourcing, not a repeat of the August 25 post.
§ 09The record
As of October 4, 2026, 00:47 UTC: page opened. Every X post above was read in full through the X API and each time computed from the post ID; view counts are as of that read. OpenAI’s surfaces were checked at 00:40 UTC; SemiAnalysis was read on its own page; OpenAI’s Path to Astra post was read on openai.com.
§ 10Sources
- @synthwavedd on X, August 25, September 2 and September 23, 2026.
- @Mr_Salio on X, September 5, 2026.
- Thibault Sottiaux on X, October 3, 2026.
- SemiAnalysis, Gemini is Cooked but GCP is Cooking, August 7, 2026, quoting its July 9 note.
- OpenAI, Path to Astra, September 1, 2026.
- Polymarket, GPT-7 released by; Manifold, GPT 7 (OpenAI) release date.
- Our records on GPT-6.1 Astra, Navier-Stokes and OpenAI Aeon.
Q1What is OpenAI Bel?
A codename for a large pretraining run that one X account, @synthwavedd, says OpenAI recently finished. The post describes it as more than 10 trillion total parameters and a possible base for models after GPT-6. OpenAI has not confirmed any of it.
Q2Who reported Bel first?
@synthwavedd posted it at 19:00 UTC on August 25, 2026. Every later write-up we found traces back to that post, including the Wccftech article and the September 5 relay that went viral.
Q3Is Bel GPT-7?
Nobody at OpenAI has said so. The source called Bel a post-GPT-6 base. A finished pretrain is a starting point that usually gets more training before release, so a product built on it could carry any name.
Q4Does GPT-6 Bel beat GPT-6 Astra?
There is no evidence either way. No benchmark, demo or comparison has been published. The ‘beats Astra’ line first appears in a September 5 relay headline, not in the original post.
Q5Is OpenAI's next-week release Bel?
Nothing links them. Thibault Sottiaux, who leads Codex, replied on October 3 that people do not know what OpenAI is releasing next week. He named no model, and the original Bel poster had said Bel was months away.
Q6Does CellCog run on OpenAI models?
No. As of October 2026 every CellCog tier runs Anthropic’s Claude Opus 5.5. We track OpenAI’s releases because many of our readers build on them.