[Video｜SemiAnalysis Founder: Anthropic's More Powerful Mythos 2 Has Been Trained but Held Back] SemiAnalysis founder Dylan revealed that Anthropic might be holding onto a model more powerful than the publicly available version, but it has not been released due to regulatory and safety pressures. He mentioned that Anthropic's Mythos had been fully trained long ago but was delayed for months, and Mythos 2 is also reportedly fully trained but remains unpublished. The publicly accessible Fable model, on the other hand, incorporates numerous safety classifiers and has very strict permission restrictions. Dylan believes the real point of concern is this: while these more powerful models are not publicly available, Anthropic is likely still using them internally to train the next generation of models. In other words, there may already be a significant gap between the AI capabilities visible to the public and those truly mastered within the lab.