Claude Mythos: Anthropic's Next Model Leaked — Everything Confirmed in April 2026
On March 26, 2026, a massive data leak exposed Anthropic's plans for its next model. Claude Mythos, the Capybara tier, unprecedented cybersecurity risks — we break down what's actually confirmed.
On March 26, 2026, a massive data leak exposed Anthropic's plans for its next model. Claude Mythos, the Capybara tier, unprecedented cybersecurity risks — we break down what's actually confirmed.
The leak that shook the AI industry
On March 26, 2026, two cybersecurity researchers — Roy Paz from LayerX Security and Alexandre Pauwels from the University of Cambridge — stumbled onto something they weren't supposed to see: nearly 3,000 internal Anthropic files sitting in a publicly accessible, unauthenticated data store due to a CMS misconfiguration.
Among those files: a draft blog post announcing a new model called Claude Mythos, described by Anthropic itself as "by far the most powerful AI model we've ever developed." Fortune reviewed the documents and notified Anthropic, which quickly locked down access. The screenshots had already spread.
Days later, a second leak: Anthropic accidentally uploaded 500,000 lines of Claude Code's original source code to npm instead of the compiled version. The two separate incidents provided overlapping evidence about a model the company had not yet intended to announce.
This guide separates what Anthropic has officially confirmed, what comes from leaked documents, and what remains unknown.
What Anthropic has officially confirmed
An Anthropic spokesperson responded to Fortune with a clear statement:
"We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. Given the strength of its capabilities, we're being deliberate about how we release it. We consider this model a step change and the most capable we've built to date."
That is everything Anthropic has officially confirmed. No benchmarks, no date, no pricing. Everything else comes from the leaked materials.
The Capybara tier: a new category above Opus
The most structurally significant detail in the leak: Mythos is not an incremental version of Opus. It is an entirely new tier in the Claude hierarchy.
Currently, Anthropic offers three model tiers:
- Haiku — fastest and most affordable
- Sonnet — the balanced model
- Opus — the flagship, most capable
Capybara and Mythos appear to refer to the same underlying model: Capybara is the tier name, Mythos is the specific model name within that tier. Two versions of the same draft blog post were found — one titled "Mythos," one "Capybara" — suggesting Anthropic hadn't finalized the commercial name.
Comparison table: Mythos vs the current Claude lineup
| Model | Tier | Status | Capabilities (per leak) |
|---|---|---|---|
| Claude Haiku 4.5 | Haiku | Available | Fast, cost-efficient |
| Claude Sonnet 4.6 | Sonnet | Available | Balanced speed/intelligence |
| Claude Opus 4.6 | Opus | Available | Best current model, 80.8% SWE-bench |
| Claude Mythos | Capybara (new) | Early access only | "Dramatically higher" than Opus 4.6 |
Presumed capabilities — what the leaked documents say
No official benchmarks have been published by Anthropic. All performance claims below come from the leaked draft, not independent testing.
Coding and academic reasoning
The leaked document states that Capybara achieves "dramatically higher scores than Claude Opus 4.6 on tests of software coding, academic reasoning, and cybersecurity." For context, Opus 4.6 already leads SWE-bench Verified at 80.8% and Terminal-Bench 2.0 at 65.4%. "Dramatically higher" without specific numbers remains unverifiable.
Cybersecurity — the most sensitive dimension
This is where the leak becomes serious. The draft describes Mythos as "currently far ahead of any other AI model in cyber capabilities" and warns that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
Anthropic doesn't hide its own concern. The document explicitly states: "In preparing to release Claude Capybara, we want to act with extra caution and understand the risks it poses." This is why the rollout begins with defensive security teams, not general users.
The risk isn't hypothetical. Anthropic had already disclosed that a Chinese state-sponsored group had used Claude Code to infiltrate roughly 30 organizations — tech companies, financial institutions, government agencies — before being detected. Mythos would significantly amplify those dual-use capabilities.
Immediate market consequences
The disclosure had measurable financial effects. The iShares Expanded Tech-Software Sector ETF (IGV) fell nearly 3% the following day. CrowdStrike, Palo Alto Networks, Zscaler, and Fortinet all declined. Bitcoin slipped toward $66,000. Anthropic simultaneously began briefing senior U.S. government officials about the national security implications of Mythos's cybersecurity capabilities.
Deployment strategy: caution as a principle
The leaked draft describes a deliberately slower rollout than previous models:
Phase 1: restricted access to a small group of early-access customers, primarily defensive cybersecurity teams. The goal is to build defensive tooling before offensive capabilities become broadly available.
Phase 2: staged expansion via the Claude API, then across Pro, Team, and Enterprise plans.
No public date. Anthropic has committed to no timeline. The document notes the model is "very expensive to serve" and that Anthropic is working to make it "much more efficient before any general release." Some analysts link a potential general release to Anthropic's IPO targeting Q4 2026, but this is speculation.
What we still don't know
To be rigorous, here is what remains unknown:
Specific benchmarks: no official figures, no independent testing. "Dramatically higher" is a qualitative descriptor, not a score.
Pricing: Anthropic has communicated nothing. The Capybara tier will necessarily cost more than Opus, but by how much — unknown.
Context window: researcher Roy Paz suggests the model will likely have "fast" and "slow" versions based on an apparently larger context window, but nothing is confirmed.
Final commercial name: "Mythos" and "Capybara" are internal names. The public name hasn't been finalized.
Release date: none. No confirmed timeline.
Claude Mythos vs the competition in 2026
Even without confirmed benchmarks, the competitive landscape helps frame the stakes.
In March 2026, Claude Opus 4.6 leads on SWE-bench Verified (80.8%) and enterprise knowledge work tasks, with a 144 Elo point advantage over GPT-5.2. GPT-5.4 (released March 5, 2026) responded with 75% on OSWorld — exceeding human performance on desktop use — and pricing 50% lower than Opus 4.6. Gemini 3.1 Pro owns the cost-efficiency segment with a 2-million-token context window at aggressive rates.
If Mythos delivers on the leaked draft's promises, it would redefine the frontier again. But by the time it reaches general availability, GPT-5.5 and Gemini 4 will also be on the market. The race doesn't stop.
Should you wait for Mythos?
The short answer: no.
Claude Opus 4.6 is Anthropic's best available model right now, and one of the most powerful on the market. Waiting for a model with no announced date, expensive to run, and initially restricted to selected cybersecurity partners makes no practical sense for the vast majority of use cases.
Workflows built today on Opus 4.6 will carry over directly when Mythos launches. Anthropic maintains careful API compatibility across generations. Starting now means being ready from day one.
Our verdict on Claude Mythos
Claude Mythos is real — Anthropic confirmed it. It's in testing with a small early-access group. The company calls it a "step change" in AI capabilities, particularly in cybersecurity. And it was revealed not through a press conference, but through two consecutive security incidents in one week — the kind of accidental disclosure that tends to generate more credible signal than staged launches.
What's certain: Mythos is the most ambitious model Anthropic has ever built. What remains uncertain: when it will be available, at what price, and whether the performance promised in the leaked documents will hold up under independent testing.
Until then, Claude Opus 4.6 remains the reference.
FAQ Claude Mythos
Is Claude Mythos officially confirmed by Anthropic?
Partially. Anthropic confirmed it is testing a new model it describes as "a step change" and "the most capable we've built to date," with "meaningful advances in reasoning, coding, and cybersecurity." But no benchmarks, pricing, or release date have been officially communicated.
What is the Capybara tier?
Capybara refers to a new model tier positioned above Opus in the Claude hierarchy — larger, more capable, and more expensive. It's a new category, not an incremental update to Opus. Mythos would be the first model in this tier.
When will Claude Mythos be available?
No date has been confirmed. The model is currently in early access for a small, selected group of customers, primarily in defensive cybersecurity. Some analysts link a potential general release to Anthropic's Q4 2026 IPO timeline, but this is speculation.
Why is Anthropic being so cautious about releasing Mythos?
Anthropic is concerned about the model's own cybersecurity capabilities. The leaked draft describes Mythos as "far ahead of any other AI model in cyber capabilities" and warns it could enable attacks that outpace defenders. Anthropic is prioritizing access for defense teams first.
Should I wait for Mythos before adopting Claude?
No. Claude Opus 4.6 is available now, performant, and compatible with future models. Waiting for Mythos means losing months of productivity with no release date guarantee.
6 articles to read next
- ChatGPT vs Claude vs Gemini: which to choose in 2026? — Chatbots, 3
- OpenAI Raises $122 Billion: $852B Valuation, SuperApp, and IPO on the Horizon — Chatbots, 20
- Is ChatGPT Losing to Claude and Gemini? What the Numbers Actually Say in 2026 — Chatbots, 11
- Anthropic Is Renting Elon Musk’s Datacenter: Why the AI War Is Becoming an Infrastructure War — Chatbots, 16
- AI Memory Changes Everything: Why ChatGPT, Claude and Gemini No Longer Want to Be Just Chatbots — Chatbots, 29
- Microsoft Copilot Review 2026: Is It Worth It for Real Work? — Chatbots, 28