Claude Mythos leaked: what Anthropic's next model means for builders
Anthropic just leaked their own next model. Not a hack. Not a whistleblower. They left 3,000 unpublished files in a public CMS cache with guessable URLs.
Among those files: two draft blog posts announcing "Claude Mythos," codenamed Capybara. A model that sits above Opus. Their most powerful ever. And it's already being tested with early access customers.
Here's what happened, what we know, and what it means if you build things with AI.
How the leak happened
On March 27, security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) found nearly 3,000 Anthropic assets sitting in a publicly searchable data store. Draft blog posts. Images. PDFs. Internal documents. All set to "public" by default.
One of those PDFs was about an invite-only CEO retreat hosted by Dario Amodei. Another was a blog post titled "Claude Mythos." Another was the same post titled "Claude Capybara." Same content, two names.
Anthropic confirmed it was human error in their CMS. No model weights leaked. No malicious breach. Just marketing materials and planning docs left in the open.
What is Claude Mythos
Mythos is a new tier above Opus. Not Opus 5. Not Opus 4.7. A whole new level. The drafts describe it as "larger and more intelligent than our Opus models, which were, until now, our most powerful."
Anthropic's model hierarchy was Opus > Sonnet > Haiku. Now it's Mythos > Opus > Sonnet > Haiku.
The name "Mythos" is meant to evoke "the deep connective tissues that link together knowledge and ideas." In plain terms: it reasons across domains better than anything they've shipped.
The benchmarks
Anthropic didn't publish exact scores in the drafts. But the language is clear: "dramatically higher scores" than Opus 4.6 in software coding, academic reasoning, and cybersecurity.
Dramatically higher than the model that already tops most coding leaderboards.
This is the part I keep thinking about. Opus 4.6 can already handle multi-file codebases that used to require senior engineers. Mythos being "dramatically" better at that means an AI agent could scaffold an entire SaaS product (auth, billing, database schema) in a single session.
With the SaaS Builder skill, I've been shipping MVPs in 3 days using Opus 4.6. If Mythos is the leap the leaks suggest, that timeline compresses to hours. The skill file becomes more valuable, not less, because a smarter model extracts more value from specific instructions.
The cybersecurity problem
Mythos is described as "currently far ahead of any other AI model in cyber capabilities" and something that "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
That's Anthropic's own language. About their own model.
Cybersecurity stocks dropped March 27-28 when this hit. Not because of the leak itself, but because of what the model can apparently do.
If Anthropic gives early access to defensive organizations, it could accelerate patching, threat detection, and secure coding. The SEO Optimizer skill already includes security checklist patterns for web apps. Imagine that level of automated security review, but for zero-day discovery across enterprise systems.
The flip side: offensive capabilities this strong, in the wrong hands, do real damage. State actors have already used Claude variants for infiltration campaigns. Mythos would make that look amateur.
What this means for builders
Putting the security angle aside for a second (and it matters enormously), think about the practical implications.
If Mythos-level coding is coming, the gap between "idea" and "shipped product" keeps shrinking. A PM with the Landing Page Builder skill and a Mythos-class model could launch a validated product over a weekend. Not a prototype. A product with users, billing, analytics.
The constraint keeps moving from "can I build this?" to "do I know what to build?" That's where skill files matter most. A model this powerful, given vague instructions, will build something impressive and wrong. Given specific architecture and patterns, it builds something impressive and correct.
The cost question
The drafts say Mythos is "very expensive for us to serve, and will be very expensive for our customers to use." Enterprise only, at least initially.
This is where skill files become a cost multiplier, not just a quality one. Every unnecessary tool call, every back-and-forth to clarify what you want, costs tokens. With Mythos pricing, that waste compounds fast. The Guerrilla Marketing skill or Taste & Design skill give the model enough context to get it right in fewer turns. Fewer turns = less cost.
When can you use it
Mythos is in early access testing now. No public release date. Anthropic is being deliberate about the rollout because of the cybersecurity risks.
Given their track record, I'd guess limited access within weeks, broader access within months. The safety review is real.
If you want to be ready when it drops, having skill files that work with today's models is the best preparation. They'll work even better with tomorrow's.
Not sure which skills match what you're building? The skill quiz takes 60 seconds.
Sources: