Anthropic did not plan to announce its most powerful AI model. A CMS misconfiguration did it for them. On March 26, 2026, roughly 3,000 internal assets — including a draft launch blog post — were left in a publicly searchable data store, forcing Anthropic to confirm what it had not yet revealed: a model called Claude Mythos, described internally as "by far the most powerful AI model we've ever developed." It represents an entirely new tier above the current Opus lineup, and Anthropic's own draft materials warn it poses "unprecedented cybersecurity risks." What those two facts mean together — for enterprises, developers, and security teams — is the real story here.
What You Need to Know
Claude Mythos is real, confirmed by Anthropic, and currently in restricted early access with select enterprise customers. It is not publicly available and has no release date. The company describes it as a "step change" above Claude Opus 4.6 in coding, reasoning, and cybersecurity — and separately warns government officials it makes large-scale cyberattacks "much more likely in 2026."
- Switch to watching this space if you work in enterprise cybersecurity, agentic AI deployment, or developer tooling — Mythos will reshape all three.
- Stay on Opus 4.6 for now if you have live production workflows — Mythos is not accessible and is explicitly described as too expensive for general release.
- Skip the 10-trillion-parameter hype if you want to make decisions — that figure is unverified and does not appear in any official Anthropic statement.
What's New: What the Leak Actually Revealed
The leak originated from a simple error. Anthropic's CMS set digital assets to public by default unless someone explicitly changed the setting to private. Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge found the exposed material before Anthropic restricted access.
An Anthropic spokesperson confirmed the model's existence, calling it a "step change" and "the most capable we've built to date," with "meaningful advances in reasoning, coding, and cybersecurity."
Here is what the confirmed and leaked materials actually state, separated from what is still speculation:
Confirmed by Anthropic
| Detail | Status | Source |
|---|---|---|
| Model exists and is named Claude Mythos | ✅ Confirmed | Anthropic spokesperson, Fortune |
| Currently in early-access testing | ✅ Confirmed | Anthropic spokesperson |
| "Step change" above Claude Opus 4.6 | ✅ Confirmed | Anthropic spokesperson |
| Poses "unprecedented cybersecurity risks" | ✅ Confirmed (leaked draft, not disputed) | Fortune |
| No general release date set | ✅ Confirmed | Anthropic |
| Too expensive to serve at scale right now | ✅ Confirmed | Anthropic |
From Leaked Draft Materials (Not Officially Confirmed)
| Detail | Status | Source |
|---|---|---|
| New "Capybara" tier above Opus | Leaked draft, acknowledged indirectly | Fortune |
| "Dramatically higher scores" in coding, reasoning, cybersecurity vs Opus 4.6 | Leaked draft | Fortune |
| Model described as "larger and smarter" than Opus | Leaked draft | Fortune |
| ~10 trillion parameters | Unverified — circulating widely, not in Fortune's reporting | Social media / speculation |
The 10-trillion-parameter figure does not appear in Fortune's reporting or in Anthropic's official statements. Treat it as unverified. It spread rapidly alongside confirmed details, but the two should not be conflated. Anthropic has not disclosed parameter counts for any of its models.
The New Capybara Tier
Anthropic currently markets models in three tiers: Opus (largest, most capable), Sonnet (faster and cheaper), and Haiku (smallest and fastest). The leaked materials describe Capybara as a new tier even larger and more capable than Opus, and more expensive.
As of early April 2026, Claude Opus 4.6 is priced at $5 input / $25 output per million tokens. The leaked draft described Mythos as "very expensive for us to serve, and will be very expensive for our customers to use," though no specific token prices were included.
How It Compares to Current Models
No independent benchmarks exist for Mythos yet. Anthropic has published no official scores, no leaderboard submissions, and no system card. What we have is Anthropic's own claim of "dramatically higher scores" vs. Opus 4.6 in three areas.
| Model | Tier | Coding | Cybersecurity | Context | Public Access | Est. Pricing (Input/Output per 1M tokens) |
|---|---|---|---|---|---|---|
| Claude Haiku 4.5 | Small | Baseline | Baseline | Standard | ✅ Yes | Lower than Sonnet |
| Claude Sonnet 4.6 | Mid | Strong | Moderate | 1M tokens | ✅ Yes | Mid-range |
| Claude Opus 4.6 | Flagship | Very strong | High (500+ CVEs found) | 1M tokens | ✅ Yes | $5 / $25 |
| Claude Mythos (Capybara) | New top tier | "Dramatically higher" | "Unprecedented" | Unknown | ❌ Early access only | Unknown — "very expensive" |
Pricing as of April 2026. Mythos pricing not yet disclosed.
The honest interpretation: Anthropic's claim of "dramatically higher scores" is self-reported and from a pre-launch draft. The company has every incentive to frame an unreleased model favorably. Independent verification is impossible until Mythos reaches third-party evaluators like Artificial Analysis or the LMSYS Arena.
The Cybersecurity Threat — and the Irony Hidden in Plain Sight
This is the part that every enterprise security team needs to read carefully.
Anthropic's own leaked draft states that "although Mythos is currently far ahead of any other AI model in cyber capabilities, it presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
According to Axios, Anthropic has been privately warning senior government officials that Mythos makes large-scale cyberattacks significantly more likely in 2026, and that agents running on systems at this capability level can plan and carry out complex operations with minimal human involvement.
The irony is difficult to overstate. Anthropic built a model so capable in cybersecurity that it is privately warning governments about it — and then accidentally exposed its details through an elementary CMS misconfiguration. The company reportedly responsible for "unprecedented cybersecurity capabilities" left 3,000 internal assets in a public data lake because a toggle was set to the wrong default.
That said, the threat narrative deserves context. In February 2026, Anthropic's Frontier Red Team published research showing that Claude Opus 4.6 — already publicly available — discovered over 500 high-severity zero-day vulnerabilities in production open-source codebases, some of which had gone undetected for decades despite expert review and millions of hours of automated fuzzing.
In other words, the cybersecurity threat from advanced AI is not a Mythos-specific problem. It is already here. Mythos, if the claims hold, would make an existing problem significantly worse — but the foundation-level risk is already embedded in current, accessible models.
A Chinese state-sponsored group ran a coordinated campaign using Claude Code to infiltrate roughly 30 organizations — including tech companies, financial institutions, and government agencies — before Anthropic detected and shut it down. The company investigated over 10 days, banned accounts, and notified affected organizations. That attack used an existing public model, not Mythos.
A Dark Reading poll published in January found that 48% of cybersecurity professionals now rank agentic AI as the top attack vector for 2026, above deepfakes and social engineering.
What Changed Since Opus 4.6
Claude Opus 4.6, released in February 2026, was already a significant model. In that same month, Anthropic launched Claude Code Security, a tool that scans codebases for vulnerabilities using reasoning rather than rule-based pattern matching — tracing data flows and catching complex flaws that static analysis misses.
Mythos, according to the leaked draft, is not an incremental update to Opus 4.6. It is a tier change — a structural expansion of Anthropic's model lineup, not a version bump. The leaked draft says Capybara is "a new name for a new tier of model: larger and more intelligent than our Opus models."
What has not changed: Anthropic still has no publicly available benchmarks for Mythos, no confirmed parameter count, no pricing, and no release date. Everything beyond the model's existence and general capability claims remains either leaked draft material or unverified speculation.
Who This Actually Affects
Enterprise security teams should act now, regardless of Mythos's release date. The threat baseline has already shifted with Opus 4.6. A single AI agent can scan for vulnerabilities and potentially exploit them faster and more persistently than hundreds of human hackers, experts warn. Waiting for Mythos to ship before updating threat models is already too late.
Developers building on the Anthropic API can safely continue on Opus 4.6 for production use. As of early April 2026, there is no public model endpoint for Claude Mythos or Capybara. The current public lineup is Haiku 4.5, Sonnet 4.6, and Opus 4.6. Mythos remains behind a closed early-access program with no public waitlist.
Enterprises in high-stakes sectors (finance, healthcare, government, critical infrastructure) need two things right now: an audit of where agentic AI tools touch sensitive systems, and a policy on "shadow AI" — employees connecting personal AI tools to internal systems. A Dark Reading poll found that 48% of cybersecurity professionals now rank agentic AI as the No. 1 attack vector for 2026.
Casual and consumer Claude users are unaffected. Mythos is not coming to Claude.ai in the near term.
Investors and SaaS companies face a different kind of exposure. The SaaSpocalypse — roughly a $2 trillion wipeout in enterprise software stocks beginning in early February 2026 — was the market processing a structural truth: if AI models can perform the tasks that enterprise software enables, the per-seat licensing model that built the SaaS industry is under pressure. Mythos accelerates that narrative.
What to Watch Next
Prediction markets currently put the odds of a public Mythos launch at roughly 73% by June 2026. That should be treated as a rough signal, not a forecast. Anthropic has said efficiency work is required before any general release — and that work is ongoing.
Three things to watch in the next 60–90 days:
- Official benchmarks: The first third-party evaluations (Artificial Analysis, LMSYS Arena) will determine whether Anthropic's "dramatically higher scores" claim holds up. Self-reported scores from pre-launch drafts are not evidence.
- Pricing announcement: Capybara-tier pricing will signal how Anthropic intends to position Mythos — broad enterprise tool or narrow specialist system for the highest-stakes use cases.
- Government response: Anthropic is privately briefing officials. Any formal policy response, export controls, or usage restrictions on frontier models would affect the entire industry, not just Mythos.
OpenAI, Google DeepMind, and xAI all have frontier releases in various stages of development. According to CNN, Mythos "presages an upcoming wave" — meaning even if Anthropic's model is currently ahead, the competitive gap is unlikely to last long.
Conclusion
Claude Mythos is real, confirmed, and not yet available. What Anthropic has established is a plausible claim of significant capability gains in coding, reasoning, and cybersecurity — and a credible, self-identified warning about what that means for the threat landscape. The 10-trillion-parameter figure circulating widely is unverified and should not drive decisions. What should drive decisions is the evidence already in front of us: current publicly available models are already autonomous enough to execute coordinated cyberattacks across 30 organizations. Mythos, if its claims hold, raises that baseline significantly. Security teams should not wait for the model to ship to act. Watch for independent benchmark results — the first third-party evaluations will be the real signal on whether Mythos is a genuine step change or a well-timed pre-launch narrative.