Anthropic says it's clearing the way for a public release of its Mythos-class models, months after holding them back over security concerns. The company had originally limited access because of what it saw as real risk to software in the wild, both public and private.

Why Mythos stayed locked down

Mythos first showed up in April as a restricted model, handed only to a short list of partners that included security researchers. Anthropic's reasoning at the time was blunt: the model was powerful enough that putting it in everyone's hands too early could hand an edge to attackers before defenders caught up.

The company framed it as a race. Whoever extracts the most value from this class of tooling wins, and in the near term that could just as easily be the people breaking into systems as the people guarding them. The longer-term bet is that defenders come out ahead, using models like this to find and patch bugs before vulnerable code ever ships.

That caution tracks with how frontier labs usually operate. Nobody ships a model this capable until the guardrails are solid enough to keep it from being turned into a weapon.

What changed

Anthropic now believes those guardrails are in place. In a blog post tied to its Opus 4.8 announcement, it confirmed Mythos-class models are heading to all customers "in the coming weeks" — though it pointedly declined to put a date on it.

The company says it's moving quickly on the remaining safeguards. Mythos is described as a real step up over Opus 4.8, its current flagship, with the biggest gains in code reasoning and autonomous operation.

The cybersecurity preview and the Claude Code slip

A small group of organizations already has hands on Claude Mythos preview for security work. Whether the exact same model is what reaches the broader public isn't clear yet.

There was also an unplanned cameo: the "Mythos-preview" model briefly surfaced for some users inside Claude Code before Anthropic pulled it back offline.