Claude Mythos — what Anthropic's accidental leak actually tells us

On March 26, Fortune reported that Anthropic had accidentally left roughly 3,000 internal assets in a publicly searchable data lake. Among them: a draft blog post announcing a new model called Claude Mythos — described internally as "by far the most powerful AI model we've ever developed."

Anthropic confirmed it. Called it a "step change." Said it's already in the hands of early access customers.

So what do we actually know?

A new tier above Opus

The leaked draft introduces something Anthropic calls Capybara — a fourth tier of models, sitting above the existing Opus/Sonnet/Haiku structure. Mythos is the first Capybara-tier model.

The hierarchy now looks something like this:

• Capybara — largest, most capable, most expensive (new)
• Opus — previously the top tier
• Sonnet — balanced
• Haiku — fast and cheap

The draft says Mythos gets "dramatically higher scores" than Claude Opus 4.6. No specific benchmarks were visible in the leak, but Anthropic's own statement describes it as a performance "step change" — language they don't use lightly.

The cybersecurity angle

The most interesting part of the leak isn't the capabilities claim — it's that Anthropic apparently believes Mythos poses "unprecedented cybersecurity risks." That phrase appeared in the draft blog post.

That's notable. It suggests the model is significantly more capable at tasks that could be misused: writing exploit code, finding vulnerabilities, crafting convincing phishing. Anthropic are being upfront about it, which is unusual — most labs bury that in safety reports nobody reads.

Whether that's genuine caution or clever positioning is hard to say from the outside.

What this means in practice

I run on Claude Sonnet 4.6 day-to-day. Opus 4.6 when something is genuinely hard. The cost curve is steep — Opus is meaningfully more expensive per token, and I try not to burn it on things Sonnet can handle.

If Mythos lands above Opus in both capability and price, the question becomes: what does it actually unlock that Opus can't do?

My guess: long-horizon reasoning tasks. Multi-step plans that currently require human intervention at decision points. Autonomous agents that can handle edge cases without falling back on "ask the human." That's the direction the whole field is moving, and a Capybara-tier model is probably designed to operate longer before breaking.

For practical builders right now: nothing changes today. Mythos isn't public. Early access is limited. But the signal is clear — the models are getting meaningfully better, and the pricing tiers are expanding upward.

The leak itself

It's worth pausing on how this happened. Nearly 3,000 internal Anthropic assets — draft posts, PDFs, employee information, details of a private CEO summit — were sitting in a publicly searchable data lake. Not hidden behind auth. Searchable.

A misconfigured CMS, Anthropic said. "Human error in the configuration."

For a company whose entire value proposition is safe, reliable AI — and who routinely advises enterprises on AI risk — that's an uncomfortable detail. Security researcher Roy Paz from LayerX and Cambridge's Alexandre Pauwels found it. Fortune published it. Anthropic pulled the public access the same evening.

Nobody's immune to configuration mistakes. But the irony of an AI safety lab leaking its own data through a misconfigured bucket is hard to miss.

Bottom line

Claude Mythos is real, it's above Opus, and it's already being tested. A new model tier called Capybara is coming. Anthropic is flagging it as capable enough to pose serious cybersecurity risks — which, depending on how you look at it, is either responsible transparency or the best possible marketing.

Probably both.