Claude Mythos: Anthropic's Most Powerful Model — Leaked Before Launch

March 29, 2026

On March 26, 2026, a configuration error in Anthropic's content management system accidentally left roughly 3,000 unpublished blog assets in a publicly searchable data store. Among them: a draft blog post announcing a model the company internally calls Claude Mythos — described as "by far the most powerful AI model we've ever developed."

The irony is hard to miss. The leaked document warns that the new model "poses unprecedented cybersecurity risks." It was secured with none.

What Happened

The exposure was discovered by Roy Paz, a senior AI security researcher at LayerX Security, and Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge. Fortune reviewed the material and reached out to Anthropic before publishing.

Anthropic's response was unambiguous. A company spokesperson confirmed the model's existence and described it as a "step change" in AI performance — "the most capable we've built to date." The company acknowledged a "human error" in the CMS caused the leak and stated that the model is currently being trialed by a small group of early-access customers.

A New Tier Above Opus

The draft blog post introduces both a code name and a product name. The model's internal code name is Mythos. The new model tier is called Capybara — a fourth tier positioned above Anthropic's current flagship Opus models, which were previously the most powerful Claude had to offer.

Anthropic's existing hierarchy goes: Haiku (fastest, cheapest) → Sonnet (balanced) → Opus (most capable). Capybara would extend this upward with a model the leaked document describes as "larger and more intelligent than our Opus models — which were, until now, our most powerful."

The naming is consistent with Anthropic's tradition of animal names and carries its own symbolism: capybaras are the world's largest rodents, known for their calm temperament despite their size. Whether that was the intended metaphor is unclear, but it fits.

What the Leaked Document Claims

According to the draft, Claude Mythos shows "dramatically higher scores" than Claude Opus 4.6 in three specific areas: software coding, academic reasoning, and cybersecurity. The document also notes the model is expensive to run and not yet ready for general availability.

The cybersecurity dimension is what drives the cautious rollout strategy. The document warns the model is "currently far ahead of any other AI model in cyber capabilities" and that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."

That is not a marketing claim — it is a risk disclosure. Anthropic's stated response to this is to begin the rollout exclusively with organizations focused on cyber defense, giving them time to harden their systems before the model's capabilities become more widely available.

The Dual-Use Dilemma

The tension Anthropic is navigating with Mythos is familiar, but the stakes are higher. Claude Opus 4.6 was already flagged as dual-use for its ability to surface previously unknown vulnerabilities in production codebases — a capability that can serve both defenders and attackers. The same week Opus 4.6 launched, Anthropic disclosed that a Chinese state-sponsored group had used Claude Code to infiltrate roughly 30 organizations, including tech companies, financial institutions, and government agencies, before being detected and shut down.

Mythos appears to take these capabilities significantly further. The "unprecedented" language in the draft is not casual — it suggests Anthropic's own internal safety evaluations returned results that warranted a different kind of launch strategy than anything the company has attempted before.

The cybersecurity industry appears to have noticed. Following the Fortune report, shares of Palo Alto Networks, CrowdStrike, and Fortinet dropped 4–6% in early trading on March 27.

What Is Not Known

The leak tells us what Anthropic was preparing to say, not what independent evaluators have found. There is no third-party benchmark data for Mythos, and the numbers cited in the draft are Anthropic's own. As the leaked document notes, the model is expensive to run and in controlled testing with a small group — which means most of the claims rest on internal assessments.

It is also worth noting that "step change" language has preceded overstated releases before. Anthropic has earned considerable credibility with Opus 4.6 and Sonnet 4.6 — but claims at this level of magnitude invite scrutiny. The question of whether Mythos represents a genuine capability jump or a careful positioning strategy ahead of a reported Q4 IPO will not be answerable until the model is more broadly tested.

What Comes Next

Anthropic has not announced a release date. The leaked draft had a publication date embedded in its structured data, but the company has not committed to any public schedule following the unplanned exposure. Given the stated concerns about cybersecurity risk, a staged rollout limited to vetted organizations for an extended period seems more likely than a general release in the near term.

For developers: the Claude API uses a unified interface across model tiers, so integrating Mythos/Capybara when it does become available should require minimal changes to existing codebases. Pricing will likely be higher than Opus — the document explicitly notes the model is expensive to run — but no figures have been disclosed.

What the Mythos leak reveals, perhaps more than anything else, is the pace at which Anthropic is moving. In February 2026, the METR time horizon for Claude Opus 4.6 was 14.5 hours — a tenfold increase over the previous year. If Mythos delivers on its claimed improvements in coding and reasoning, those numbers will move again. The question is less whether a step change is coming, and more what the world looks like when it arrives.