OpenAI's 10T-parameter 'Bel' pretrain points to GPT-6
OpenAI just finished pretraining "Bel," a massive foundation model packing over 10 trillion parameters, according to insider reports. The model is slated to serve as the base for Astra and GPT-6 after reinforcement learning, with sources claiming it could reach an AGI threshold—while Anthropic reportedly braces for a quiet stretch amid severe compute constraints.
A new leak suggests OpenAI just wrapped up training its next massive foundation model.
A 10T parameter beast named Bel
According to a scoop shared by @synthwavedd on X, OpenAI recently completed the pretraining run for a model codenamed Bel, the direct successor to Doug.
Bel reportedly packs over 10 trillion total parameters, putting it in the same weight class as GPT-4.5. With further reinforcement learning, it is expected to serve as the base for Astra and GPT-6, and potentially even OpenAI's target model for hitting the AGI threshold.
The leak notes that Anthropic is feeling the heat. OpenAI believes Anthropic currently lacks a strong public counter to Astra, mostly due to compute constraints. Anthropic is reportedly bracing for OpenAI to dominate the narrative through the rest of the year, though they expect to bounce back early next year.
Why it matters
The scaling race is clearly far from dead. If Bel delivers on its promises, users and developers could see a substantial leap in frontier reasoning when Astra and GPT-6 finally arrive.
It also highlights how brutal compute bottlenecks have become at the cutting edge. Even a top-tier lab like Anthropic can get temporarily boxed out simply because it lacks the raw chips to train and deploy a counter punch in time.
We think OpenAI's strategy of sheer brute-force scale will keep it ahead through year-end, but training a 10-trillion-parameter monster is one thing—making the inference economics work for real-world products is a whole different battle.
Sources (1)
- x.com(opens in a new tab)Originalx.com