Claude Mythos: Anthropic Confirms an Above-Opus Model Class

Everyone talks about AI models like they're on a ladder — each one a little better than the last. Nobody tells you what happens when the company building the ladder says the next rung isn't a step, it's a floor change.

TL;DR: Anthropic confirmed it's developing a model class above Claude Opus 4.6, internally referred to as "Mythos" or "Capybara." Existence is confirmed by Anthropic's own spokesperson. Capability claims — dramatically higher reasoning, coding, and cybersecurity scores — come from a leaked internal draft, not published benchmarks. No release date. Rollout starts with a small cybersecurity early-access group.

What Actually Happened

A data breach at Anthropic — separate from the Claude Code source leak making headlines this week — exposed internal draft blog posts describing an unreleased model. Anthropic confirmed to Fortune that the model exists and that it represents what they called a "step change" above the current Opus line.

Two name candidates appeared in the leaked drafts: Mythos and Capybara. The documents describe dramatically higher benchmark scores across reasoning, coding, and cybersecurity evaluations compared to Claude Opus 4.6. Here's the thing — those numbers come from an internal draft that was never meant to be published. No independent benchmark confirmation exists. Nobody outside Anthropic has tested this model.

So what we actually know: the model is real. Anthropic said so. What it can do? That's still their word against a leaked document.

Why This Matters

Anthropic's current top model — Claude Opus 4.6 — is already what powers the most capable AI tools available to individual users and businesses right now. If the next model is genuinely a class above that, it changes the math on what AI can handle autonomously.

But there's a reason Anthropic is rolling this out slowly, starting with cybersecurity-focused early-access customers. The leaked drafts describe the model as "very expensive" to serve, and the company is tying general availability to cost reduction work. Translation: even if Mythos performs as described, most people won't have access for a while.

The cybersecurity-first rollout is also telling. When a company leads with security researchers instead of developers or consumers, they're signaling they think the model's capabilities carry real risk — and they want safety-focused users stress-testing it first.

What To Do About It

• Don't plan around capabilities that aren't confirmed. The leaked benchmarks are interesting, but they're not published, peer-reviewed, or independently verified. Treat them as signal, not fact.
• Watch for the early-access announcement. When Anthropic opens the cybersecurity preview, that's when real performance data starts flowing from people who've actually used it.
• Check your current model. If you're building on Claude Opus 4.6 today, nothing changes right now. A model that doesn't exist in production yet doesn't affect your stack.

Now you know what Anthropic confirmed — and where the line sits between confirmed and speculation. When Mythos actually ships, that's a different conversation entirely.

Key Takeaways

• Anthropic confirmed to Fortune that it is developing a model class above Claude Opus 4.6, described as a "step change."
• Two internal name candidates exist: Mythos and Capybara. Neither is a confirmed product name.
• All capability claims — including dramatic benchmark improvements — originate from a leaked internal draft, not published or independently verified benchmarks.
• Initial rollout targets cybersecurity early-access customers, suggesting Anthropic views the model's capabilities as carrying dual-use risk.
• No release date or general availability timeline has been announced.