FULL STORY

OpenAI's Next-Gen Model Rumors: From Astra to the 10T-Parameter Bel

Rumors surfaced in late August that OpenAI is testing a far more capable next-gen model codenamed Astra, followed by Reddit leaks claiming a 10-trillion-parameter model, Bel, has finished pretraining. OpenAI has not confirmed either.

2026-08-24 ~ 2026-08-27 · 2 episodes · 16 posts

Episode 1 · OpenAI's next model Astra reportedly solves hard problems, release delayed by risk (2026-08-24, 3 posts)

Multiple reports claim OpenAI's next model Astra has solved long-standing research problems but was delayed after hitting the company's highest cyber risk threshold. A 10T-parameter preview model 'bel' and Anthropic's next-gen Mythos are also rumored to be in testing.

Episode 2 · Rumors: OpenAI completes 10T-parameter pretraining of model 'Bel' (2026-08-26, 13 posts)

Starting August 26, multiple Reddit leakers (Leo's claim relayed by Outside-Iron-8242, plus ChrisGPT, DrSingularity, scaling01, VraserX, TheDigitalHearts, and reposts by ivanbezdomny, haider1, rohanpaulai, imjustnewatai, koltregaskes) posted highly overlapping claims: OpenAI has completed a major pretraining run codenamed Bel with over 10T total parameters, potentially the foundation of its next products. OpenAI has made no response; all information is third-party rumor, and repetition does not equal verification.

Unconfirmed

  • Existence and codename: ChrisGPT, self-described as the first leaker, says OpenAI's next major pretraining is codenamed Bel, a 'killer' monster model aimed at outdoing Google's Gemini line; per Leo via Outside-Iron-8242, pretraining is complete.
  • Scale: Leakers uniformly cite >10T parameters; scaling01 adds it is roughly the size of GPT-4.5 and that OpenAI internally believes it may even reach the AGI threshold.
  • Lineage: scaling01 and DrSingularity call Bel the successor to 'Doug', serving as the base for Astra and GPT-6 (after further RL). Per koltregaskes' relayed sources, after Astra—expected in September, possibly delayed by regulation—Bel is slated for 2026 to surpass 'Fable', described as a 'giant squid'.
  • Timing: ChrisGPT expects Bel by year-end or a few months after Astra; koltregaskes' sources point to 2026—accounts conflict.
  • Training mechanism: imjustnewatai speculates Bel will never be 'done' training, using fast/slow dual-speed continual learning—fast layers absorbing verified proofs, code and tool traces, slow cycles iterating separately; this is the blogger's personal guess.

Why it matters

  • If true, Bel would be OpenAI's key base model for the GPT-6/Astra era, directly shaping its race with Google's Gemini line; rohanpaulai and ivanbezdomny note that compute-constrained Anthropic could stay passive all year.
  • Per 新智元, Anthropic plans to release Opus 5.1 this week, boosting per-token intelligence, coding, and long-running agent capability—possibly a response; WorldofAI's roundup also cites Anthropic quietly testing a Claude codenamed Marshmallow and GLM-5.3-Flash appearing.
  • VraserX reads this as 'the start of recursive self-improvement' and the singularity—personal speculation.
  • Crucially, the leakers' accounts overlap heavily but no primary evidence exists; readers should stay cautious.