The successor to OpenAI's Doug project provides the foundational layer for the upcoming Astra and GPT-6 systems. Codenamed Bel, the pretrain contains more than 10 trillion parameters, a scale comparable to GPT-4.5 that the company believes could support a model reaching an artificial general intelligence threshold. Bel is intended as a post-GPT-6 base, serving as the core for the next series of frontier models.

Anthropic lacks the compute to prepare a public response to Astra this year, forcing the company to brace for OpenAI's dominance for the remainder of 2026. The competitor expects to return to the lead in early 2027 by leveraging compute capabilities from Project Stargate to train more advanced models. OpenAI believes Anthropic's current constraints prevent any immediate challenge to the Astra launch.

Sign in to suggest edits

Key sources

  1. SOURCE@synthwavedd“potentially even the base for an AGI-threshold model”x.com
  2. SUPPORT@nicdunz“similar in size to GPT-4.5”x.com
  3. SUPPORT@kimmonismus“Project Stargate is apparently paying off, but not in the way expected”x.com
Markdown