Anthropic has quietly started rolling out Claude Mythos (internally codenamed Capybara) in stealth mode. This is not just another update—it's a next-generation model that, according to the company itself, represents a "step change" in AI capabilities.
What is known:
— Mythos has already appeared in Claude Code as an experimental option (see screenshot)
— According to internal benchmarks, the model scores "dramatically higher" compared to Claude Opus 4.6 in coding, academic reasoning, and cybersecurity
— Anthropic itself calls Mythos "the most capable model we have built to date"
— The model is being tested with a narrow group of early access clients—it's expensive to launch and not yet ready for mass release
The existence of Mythos became known to the world by accident—about 3,000 unpublished documents, including a draft blog post, ended up in a public cache due to a CMS error. Fortune managed to read everything before Anthropic closed access.
An amazing fact: Anthropic has outpaced OpenAI in launching a next-generation model. Just a year ago, this was unthinkable—Anthropic was perceived as always playing catch-up. And now they are the first to roll out next-gen, while OpenAI is still refining its current models.
