Key Notes
- The claim: OpenAI finished pretraining a 10-trillion-plus-parameter base model codenamed Bel, intended to underpin its Astra project and eventual GPT-6, and possibly an "AGI-threshold" system after further reinforcement learning; none of this is independently verifiable.
- A separate, more serious claim in the same rumor chain says OpenAI "believes" Anthropic has no viable answer to Astra due to compute constraints.
An unverified claim circulating on X since August 25 says OpenAI has completed pretraining a new base model codenamed “Bel,” reportedly exceeding 10 trillion parameters, which would position it as the foundation for OpenAI’s in-development Astra system and, eventually, a future GPT-6 generation. The claim originated from a single anonymous account and was subsequently amplified by other social media users and several aggregator sites. OpenAI has not confirmed the existence of Bel, any related codenames, or the completion of any specific pretraining run.
🚨 SCOOP: OpenAI recently finished its next pretrain, codename “Bel” – the successor to “Doug”, which is expected to be the base for Astra and GPT-6 (w/ further RL)
Bel is a giant pretrain with >10T total parameters, similar in size to GPT-4.5. OpenAI expect it to be a…
— leo 🐾 (@synthwavedd) August 25, 2026
It is worth being explicit about the reliability problem here. This is an anonymous, single-source social media claim with no corroborating documentation, and even the secondary accounts repeating it disagree on basic details, including whether Bel succeeds or precedes another rumored model codenamed “Doug.” Parameter counts, training methodology and timelines in this class of rumor have historically proven unreliable across the industry, and nothing here should be read as confirmed fact.
What is separately documented is Astra itself, which OpenAI has publicly discussed. The company has described Astra as a new class of model built around a coordination layer that can break a complex problem into pieces, assign sub-agents to work on them, and synthesize the results, sometimes running for hours or days on a single task. In early August, OpenAI said Astra had independently solved or made substantial progress on ten longstanding problems in mathematics and theoretical computer science, with solutions verified using the formal proof language Lean 4, a claim OpenAI itself made and published.
Why the Anthropic Claim Deserves Extra Scrutiny
The most consequential part of this rumor is also its least substantiated: a claim that OpenAI “reportedly believes” Anthropic has no viable competitive response to Astra because of chronic compute constraints, and that this could pressure Anthropic’s planned IPO. This assertion traces back to the same anonymous social media chain, offers no sourcing beyond that, and should not be treated as a credible report on Anthropic’s internal technical position.
It is also difficult to square with what has actually been disclosed. Anthropic reported an annualized revenue run rate reaching $65 billion by late July, has struck major compute agreements with Amazon, Google and AMD this year, and has been preparing investor meetings for a potential IPO valued at $2 trillion or more, according to multiple corroborated reports from Bloomberg, Reuters and the Financial Times.
None of that guarantees Anthropic has compute parity with OpenAI on any specific frontier training run, a genuinely contested and opaque question industry-wide, but an anonymous claim asserting Anthropic has “no viable contender” is speculation dressed as insider knowledge, not a verified fact.
The Broader Pattern
This episode fits a recurring pattern in AI coverage, where unverified leaks about next-generation models circulate rapidly, get repeated by aggregator sites as though sourced, and shape market and public perception before any company confirms or denies them.
Model codenames, alleged parameter counts and competitive claims of this kind should be treated as speculation until an AI lab makes an official statement, and readers should be skeptical of any framing that presents one company’s internal beliefs about a rival’s weaknesses as established fact, particularly when the only source is an anonymous social media post.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
