Claude Mythos Capybara: Anthropic’s Most Powerful AI Model Explained
Claude Mythos Capybara is Anthropic’s most advanced AI model, representing what the company calls a “step change” in artificial intelligence capability. The model was first reported by Fortune on March 26, 2026, after a CMS misconfiguration exposed roughly 3,000 unpublished internal assets. Capybara is not a version update — it is an entirely new tier in Anthropic’s model hierarchy, sitting above Opus as the highest class of Claude model ever built.
What Is Claude Mythos Capybara
The name “Claude Mythos” refers to a specific model, while “Capybara” refers to the new tier it belongs to. Anthropic’s leaked draft blog post stated plainly: “Capybara is a new name for a new tier of model: larger and more intelligent than our Opus models — which were, until now, our most powerful.” This distinction matters because previous Claude updates (like moving from Opus 4 to Opus 4.6) were incremental improvements within a tier. Capybara breaks that pattern entirely.
The Capybara Tier Explained
Anthropic’s model lineup has historically followed a three-tier structure: Haiku for lightweight tasks, Sonnet for balanced performance, and Opus as the flagship. Each tier serves a different cost-performance trade-off. Capybara adds a fourth tier above Opus, designed for problems that require breakthrough-level reasoning, deep cross-domain knowledge synthesis, and capabilities that Opus cannot match.
The naming convention follows Anthropic’s pattern of using animal-related terms — capybaras being the largest living rodents, fitting for the largest model class. The full hierarchy now reads: Haiku, Sonnet, Opus, Capybara — from fastest and cheapest to most powerful and most expensive.
Mythos as the First Capybara Model
Claude Mythos is the inaugural model in the Capybara tier. Internal documents revealed two candidate names — “Mythos (v1)” and “Capybara (v2)” — suggesting Anthropic was still finalizing branding when the leak occurred. The company has since confirmed Mythos as the model name and Capybara as the tier designation.
Anthropic’s spokesperson described it as “the most capable we’ve built to date” and a “general purpose model with meaningful advances in reasoning, coding, and cybersecurity.” This is not a narrow specialist — Mythos is designed to be the best at everything Claude does, just at a higher performance ceiling and a higher operational cost.
How It Compares to Claude Opus 4.6
The leaked benchmarks paint a clear picture. Compared to Claude Opus 4.6, Mythos achieves “dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity.” While Anthropic has not published exact numbers, the language in the draft blog post — “dramatically” rather than “incrementally” or “modestly” — signals a substantial performance gap.
A community-sourced comparison gives a sense of where each tier sits:
| Dimension | Haiku 4.5 | Sonnet 4.6 | Opus 4.6 | Capybara (Mythos) |
|---|---|---|---|---|
| Speed | Fastest | Fast | Medium | Slowest (expected) |
| Reasoning | Basic | Good | Excellent | Breakthrough |
| Cybersecurity | Limited | Moderate | Strong | “Far exceeds peers” |
| Cost | Lowest | Medium | High | Highest |
The trade-off is clear: Capybara delivers superior intelligence at the expense of speed and cost. Anthropic acknowledged this directly, stating the model is “very expensive for us to serve, and will be very expensive for our customers to use.”
How Claude Mythos Was Leaked
The exposure of Claude Mythos was not a sophisticated cyberattack. It was a configuration error — the kind of mundane mistake that cybersecurity professionals warn about constantly, and deeply ironic given the model’s own cybersecurity focus.
The CMS Misconfiguration
Anthropic’s content management system had a default setting that made uploaded assets publicly accessible unless explicitly marked as private. Approximately 3,000 unpublished digital assets, including draft blog posts, internal announcements, and marketing materials about Mythos, ended up in an unencrypted, publicly searchable data store. Anyone who knew where to look could access them.
Anthropic attributed the exposure to “human error in the configuration of its CMS tools.” The misconfiguration likely went undetected for some time before independent researchers found the data.
Who Discovered the Leak
Two security researchers independently identified the exposed assets. Roy Paz from LayerX Security and Alexandre Pauwels from the University of Cambridge both discovered the publicly accessible cache and reported it. Fortune was the first major outlet to publish a story based on the leaked materials, contacting Anthropic for comment before running the exclusive.
The irony was not lost on observers: Anthropic’s draft blog post about Mythos emphasized the model’s “unprecedented cybersecurity risks” while the company itself had failed to secure its own content management infrastructure. Futurism’s headline captured the sentiment: “Anthropic Just Leaked Upcoming Model With ‘Unprecedented Cybersecurity Risks’ in the Most Ironic Way Possible.”
Anthropic’s Official Response
Rather than denying or minimizing, Anthropic confirmed the model’s existence within hours. The company stated it was testing Mythos with a select group of early-access customers and described the model as representing a step change in capability. Anthropic emphasized its “deliberately slow” rollout strategy, framing the restricted access as a safety measure rather than a commercial delay.
Claude Mythos Capybara Capabilities
The leaked draft materials and Anthropic’s subsequent confirmation outline three primary areas where Mythos outperforms all existing AI models, including its own predecessor.
Coding and Software Development
Mythos scores dramatically higher than Opus 4.6 on software coding benchmarks. This continues the trajectory that established Claude Code as one of the most capable AI-assisted development tools available. The improvement suggests Mythos can handle more complex codebases, longer contexts, and more nuanced architectural decisions than any previous Claude model.
For developers, this translates to better code generation, more accurate bug detection, and improved ability to refactor large-scale systems. Apiyi’s analysis lists “large codebase refactoring” as a primary Capybara use case — tasks that require understanding thousands of interconnected files simultaneously.
Academic Reasoning Performance
The model shows “significantly improved” academic reasoning, indicating stronger performance on structured multi-step thinking tasks. This includes mathematical proofs, scientific analysis, and cross-domain knowledge synthesis that Opus 4.6 handles well but not flawlessly.
Anthropic’s internal description mentions the model’s ability to create “deep connective tissue between ideas and knowledge” — suggesting Mythos excels not just at answering individual questions but at identifying relationships across disparate fields. This capability makes it particularly useful for research, academic writing, and complex analytical tasks.
Cybersecurity: Dual-Use Concern
This is the capability that generated the most headlines. Anthropic’s draft stated that Mythos is “currently far ahead of any other AI model in cyber capabilities” and warned it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”
The model demonstrated an ability to surface previously unknown vulnerabilities in production codebases — a capability Anthropic acknowledged as dual-use. In the right hands, it strengthens defense. In the wrong hands, it enables offense at scale. This concern is not theoretical: Anthropic’s own leaked materials referenced a previous incident where a Chinese state-sponsored group used Claude Code’s agentic capabilities to infiltrate approximately 30 organizations worldwide.
Market Impact of the Capybara Leak
The financial markets responded quickly and negatively to the Mythos revelation, particularly in the cybersecurity sector.
Cybersecurity Stock Declines
Within 24 hours of the leak, several major cybersecurity companies saw their stock prices drop sharply. Palo Alto Networks (PANW) fell approximately 7%, while CrowdStrike (CRWD) declined roughly 6.4%. Zscaler (ZS) dropped about 5.8%, and Fortinet (FTNT) slipped around 4%.
The reasoning was straightforward: if an AI model can discover and exploit vulnerabilities faster than human security teams can patch them, the value proposition of traditional cybersecurity platforms faces a structural challenge. As Scott Dylan of Nexatech Ventures noted, investors were “pricing in the structural risk that foundation model providers can now compete directly with the software layer.”
Anthropic IPO Implications
The leak arrived at a sensitive time for Anthropic. Bloomberg and The Information have reported the company is considering an IPO in October 2026 with a projected valuation of $60 billion or more. A model that demonstrably outperforms everything on the market could boost that valuation — but the circumstances of its revelation (via a security blunder) and the regulatory concerns around its cybersecurity capabilities complicate the narrative.
The Capybara tier, if priced at a premium above Opus, also represents a significant new revenue stream. Community estimates suggest Capybara API pricing could reach $10-25 per million input tokens and $50-125 per million output tokens — two to five times more expensive than Opus 4.6’s current $5/$25 pricing.
When Will Capybara Be Available
Current Access: Restricted Testing
As of March 28, 2026, Claude Mythos is not publicly available in any form. Anthropic has confirmed that the model is being tested with a small group of early-access customers focused specifically on cybersecurity defense applications. The goal is to give defenders a head start in hardening their systems before broader deployment.
Expected Public Release Timeline
Anthropic has not announced a public release timeline date. The company has explicitly stated that release timing will be “determined by safety evaluations, not commercial schedule.” However, several factors point to a likely window.
Industry analysts predict expanded testing in Q2-Q3 2026, with a public launch most likely in late 2026 — potentially coinciding with the anticipated IPO in October 2026. The commercial incentive to launch before going public is substantial, as a live Capybara product would strengthen the IPO pitch.
Estimated Pricing
Anthropic has not released official pricing. The leaked materials confirm that the model is expensive to run, and the company is working to improve efficiency before general release. For reference, the current Opus 4.6 pricing sits at $5 per million input tokens and $25 per million output tokens. Capybara will almost certainly cost more — the question is how much more.
Given the four-tier structure and Anthropic’s track record of roughly 2-3x pricing jumps between tiers (Haiku at $0.80/$4, Sonnet at $3/$15, Opus at $5/$25), a reasonable estimate places Capybara in the $10-15 per million input tokens and $50-75 per million output tokens range, though some community projections go higher.
Questions About Claude Mythos Capybara
What is Claude Mythos?
Claude Mythos is Anthropic’s most powerful AI model, leaked on March 27, 2026 through a CMS misconfiguration. It is the first model in the new Capybara tier, which sits above Opus in Anthropic’s model hierarchy.
What does Capybara mean in Anthropic’s model lineup?
Capybara is the name for Anthropic’s new highest model tier. The hierarchy is Haiku (lightweight), Sonnet (balanced), Opus (flagship), and Capybara (breakthrough). Mythos is the first — and currently only — Capybara-tier model.
Is Claude Mythos available to the public?
No. As of March 2026, Mythos is restricted to a small group of early-access customers testing cybersecurity defense applications. There is no public API access or release date announced.
How was Claude Mythos leaked?
A misconfiguration in Anthropic’s content management system left approximately 3,000 unpublished assets in a publicly searchable, unencrypted data store. Security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) independently discovered the exposure.
What are Claude Mythos benchmark scores?
Anthropic has not published exact numbers. The leaked draft states Mythos achieves “dramatically higher scores” than Claude Opus 4.6 on software coding, academic reasoning, and cybersecurity tests. The word “dramatically” rather than “modestly” suggests a significant gap.
How much will Capybara tier cost?
No official pricing exists yet. Current Opus 4.6 costs $5/$25 per million tokens (input/output). Community estimates for Capybara range from $10-25 input and $50-125 output per million tokens, representing a 2-5x premium over Opus.
When will Anthropic release Capybara?
No confirmed date. Anthropic says release timing depends on safety evaluations. Analysts predict expanded testing in Q2-Q3 2026 and a possible public launch in late 2026, potentially aligned with Anthropic’s anticipated October 2026 IPO.
What is the difference between Capybara and Opus?
Capybara is a tier above Opus — larger, more intelligent, and more expensive. While Opus excels at complex reasoning and research, Capybara achieves breakthrough-level performance in coding, reasoning, and cybersecurity. The trade-off is slower speed and significantly higher cost.