Anthropic's Secret Model 2 Beats Mythos 5. But there is a catch, you Can't Use It

Anthropic just confirmed it: there's a secret AI model inside the company, called Model 2, and it's better than Claude Mythos 5, the company's top public model right now. And no, you can't try it.

The news showed up in a place most people would skip: Anthropic's August 2026 Risk Report, a routine safety paper that reads more like homework than headline news. But buried inside is proof that Anthropic's next big model already exists and is being used inside the company right now.

On Anthropic's own test, called CoBench, Model 2 scores 62.8%. Mythos 5 only hits 50.3%. That's a solid win. But Anthropic is quick to cool things down, calling it a nice step forward, not some huge leap like the jump from Opus 4.6 to Mythos Preview.

It's not because Model 2 failed some scary safety test. Anthropic says the model passed its checks and showed no new warning signs beyond what we already know about Mythos 5. The real reason is simpler: it just hasn't finished all of Anthropic's usual pre-release testing yet. So the company doesn't feel fully sure about its safety picture, and isn't ready to hand it over. No release date is set.

Anthropic's next flagship model is already alive and working behind closed doors. This is the same pattern we saw with Mythos Preview before it went public. Will "Model 2" become Mythos 6? A rebrand? Something brand new entirely? Nobody knows yet. What we do know: it's real, it's stronger, and it's not coming out anytime soon.

Put simply, Anthropic already built something better than its best public model, and it's sitting there, unreleased, for now. This doesn't mean a launch is right around the corner. But it does mean the future is already here, just not ready to meet the public yet.