Anthropic has confirmed: they are testing a new model at a "step change" level.

Code name Mythos, internally called Capybara. The leak was accidental, but the model is real.
What did they say about previous AI releases? "Improved by X% on the XX benchmark," "faster and cheaper," "supports longer context."
And what about Mythos? "Far surpasses any other AI model in cybersecurity capabilities," "signaling a wave of models that can exploit vulnerabilities in ways far beyond defenders."
Can you see the difference?
Before, it was "we are better." This time, it's "we are so powerful that it could pose a risk."
Anthropic has always been the company with the strongest AI safety narrative. Their RSP (Responsible Scaling Policy) essentially sets limits for themselves—if a model's capabilities exceed a certain threshold, additional safety measures must be implemented before release.
Mythos might be the first model to truly trigger a high-level safety review.
In the short term, Anthropic may be very cautious. "Limited early access for select clients" + "focusing on cybersecurity defenders." Not a full consumer release.
In the long run, the AI capability curve may not be linear. "Step change" means a leap—going from the previous model to this one is not just a 10% improvement, but a qualitative leap.