OpenAI’s Bel Leak Claims 10T Parameters

The interesting question is not whether Bel crosses 10T parameters, but whether OpenAI can turn that scale into reliable, affordable intelligence. For now, this is a compelling rumor—not a model release.

• –“Total parameters” may describe a mixture-of-experts system; active parameters will matter more for inference cost and latency.
• –Completing pretraining says little about post-training, evaluations, safety, or product readiness.
• –A shared base for future reinforcement-learning systems would reinforce the industry’s shift toward separating pretraining from capability shaping.
• –Developers should wait for verified benchmarks, architecture details, and API access before drawing conclusions from the parameter count.
• –Larger models can require more inference compute, making efficiency and serving infrastructure as important as raw scale. [OpenAI research](https://openai.com/index/evaluating-chain-of-thought-monitorability/)