Claude Mythos (Capybara): Everything Leaked About Anthropic's Most Powerful AI Model
Anthropic accidentally exposed 3,000 internal files revealing a new AI model called Claude Mythos, codenamed Capybara. It beats Opus 4.6 on every benchmark, introduces a new model tier above Opus, and poses what Anthropic itself calls unprecedented cybersecurity risks. Here is everything confirmed.
TL;DR
An unreleased Anthropic model accidentally confirmed via a CMS data leak on March 26, 2026.
Capybara sits above Opus in a brand new fourth tier, making it the most powerful Claude model ever built.
Dramatically higher scores than Opus 4.6 on coding, academic reasoning, and cybersecurity benchmarks.
Anthropic is holding it back due to unprecedented cybersecurity risks and very high serving costs.
Limited early access for cybersecurity defense customers only. No public release date confirmed.
Claude Mythos (Capybara): Everything Leaked About Anthropic's Most Powerful AI Model
Anthropic did not plan to tell the world about its most powerful AI model this week. A misconfigured content management system did it for them.
On March 26, 2026, security researchers discovered that Anthropic had accidentally left close to 3,000 unpublished internal assets in an unencrypted, publicly searchable data cache. Among them was a draft blog post describing a new model called Claude Mythos, also referred to internally as Capybara, which the company describes as a step change in AI capability and the most powerful model it has ever built.
Within hours of Fortune reporting the leak, Anthropic confirmed the model is real, cybersecurity stocks fell 4 to 6 percent, and the AI community was processing what the disclosure actually means. Here is a complete breakdown of everything that is confirmed, what remains unknown, and what this model could change.
How the Leak Happened
The incident was discovered independently by two security researchers: Roy Paz, a senior AI security researcher at LayerX Security, and Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge.
The leak was caused by a configuration error in Anthropic's content management system, which made close to 3,000 unpublished assets publicly accessible. Anthropic attributed the incident to "human error" in the configuration of its CMS, describing the exposed material as "early drafts of content considered for publication."
Digital assets including images, PDF files, and audio files were set to public by default upon upload, unless explicitly marked private. This oversight led to approximately 3,000 assets linked to Anthropic's blog becoming publicly accessible.
Fortune reviewed the documents and informed Anthropic on Thursday evening. Anthropic then restricted public access to the cache. The company acknowledged the model's existence in a statement to Fortune the same night.
The exposed files included not only the model announcement drafts but also details of a planned invite-only CEO summit at an 18th-century English countryside manor, which Anthropic CEO Dario Amodei is set to attend. The event is described as part of an ongoing series aimed at selling Anthropic's models to large European corporate customers.
What Is Claude Mythos?
Claude Mythos is Anthropic's next flagship AI model, currently in restricted testing. It exists under two names: Mythos is the model name, while Capybara appears to be the name of the new tier it belongs to. Two versions of the same draft blog post differ only in the model's name: Mythos in version one and Capybara in version two. The Capybara version swaps the name throughout the title and body text, but the subtitle still reads "We have finished training a new AI model: Claude Mythos." Both versions use the same justification for the name, saying it was chosen to evoke "the deep connective tissue that links together knowledge and ideas."
Anthropic's spokesperson confirmed the model in a 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. As is standard practice across the industry, we're working with a small group of early access customers to test the model. We consider this model a step change and the most capable we've built to date."
A New Model Tier Above Opus
Claude Mythos does not slot into Anthropic's existing model hierarchy. It introduces an entirely new tier.
Anthropic currently markets models in three tiers: Haiku (smallest, fastest, cheapest), Sonnet (balanced), and Opus (largest, most capable, most expensive). Capybara would add a fourth, pricier tier above all three. According to the draft, it scores "dramatically higher" than Claude Opus 4.6 on tests of software coding, academic reasoning, and cybersecurity.
The leaked draft makes the positioning explicit: "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 is a structural change to Anthropic's model lineup, not a version increment within an existing tier.
What Claude Mythos Can Do
The leaked documents describe three areas of exceptional performance.
Software coding: The model scores dramatically higher than Claude Opus 4.6 on coding benchmarks. For context, Opus 4.6 recently topped Terminal-Bench 2.0 at 65.4 percent, surpassing GPT-5.2-Codex. Mythos's score on the same benchmark was not specified in the leaked draft, but the "dramatically higher" characterization suggests a substantial gap.
Academic reasoning: The model outperforms Opus 4.6 on academic reasoning tests. No specific benchmarks or scores were included in the leaked materials.
Cybersecurity: This is where the model is both most capable and most concerning. Anthropic says the model is "currently far ahead of any other AI model in cyber capabilities" and "it presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
The leaked drafts show that the model achieves significantly higher scores than Claude Opus 4.6 in tests involving software programming, academic reasoning, and cybersecurity, with cybersecurity performance far surpassing all existing AI models.
Why Anthropic Is Withholding It
The cybersecurity performance is the primary reason Claude Mythos is not publicly available. The leaked draft is explicit about the concern.
"In preparing to release Claude Capybara, we want to act with extra caution and understand the risks it poses, even beyond what we learn in our own testing. In particular, we want to understand the model's potential near-term risks in the realm of cybersecurity, and share the results to help cyber defenders prepare," the document said.
The dual-use problem is real. The same capabilities that make Mythos valuable for identifying vulnerabilities in defensive security contexts also make it potentially dangerous in the hands of malicious actors. A model that can find and exploit software vulnerabilities faster than defenders can patch them does not just help the security teams who deploy it.
This is not hypothetical. Anthropic has previously disclosed that a Chinese state-sponsored group had already been running a coordinated campaign using Claude Code to infiltrate roughly 30 organizations including tech companies, financial institutions, and government agencies before the company detected it. Over the following 10 days, Anthropic investigated the full scope of the operation, banned the accounts involved, and notified affected organizations.
A model that is, by the company's own description, far ahead of every other AI in cybersecurity capabilities raises that threat profile considerably.
The rollout will start with a small group of early-access customers tasked with evaluating cybersecurity applications under an initiative heavily rumored to be called Project Glasswing. This project focuses on allowing critical cyber-defense organizations to identify and patch security vulnerabilities before the model ever sees wider availability. The drafts also acknowledge the model is "very expensive for us to serve, and will be very expensive for our customers to use." To track when this pricing is finalized, keep an eye on our AI Tool Directory.
The Market Reaction
The leak moved markets within hours of Fortune's report going live.
Among the names moving sharply lower on the news: Palo Alto Networks, CrowdStrike, and Fortinet all fell 4 to 6 percent. The broader iShares Expanded Tech-Software Sector ETF fell approximately 2.5 percent.
The reasoning from investors was straightforward: a model with superior vulnerability detection undermines the value proposition of existing cybersecurity products. If an AI can find and explain software vulnerabilities faster and more comprehensively than current security scanning tools, the competitive position of those tools weakens.
Whether that concern is proportionate remains to be seen. The model is not publicly available, and Anthropic has committed to a cautious release strategy focused initially on cybersecurity defense rather than offensive use.
The Irony of the Leak Itself
The timing was not lost on observers. In an enormously ironic twist, a draft blog obtained by Fortune, which was available in an unsecured and publicly searchable data store, claimed that the new model "poses unprecedented cybersecurity risks." In other words, a company warning about cybersecurity dangers from its own AI model had its own basic content management security misconfigured.
Anthropic acknowledged the error. The company described it as human error in CMS configuration rather than a malicious breach, and access was restricted the same evening Fortune made contact.
What We Do Not Know
The leaked materials are draft documents, not a finished announcement. Several important details remain genuinely unknown.
Exact benchmark scores. The draft uses the phrase "dramatically higher" throughout but does not include specific numbers. Until Anthropic publishes official evaluation results, there is no way to quantify how large the gap between Mythos and Opus 4.6 actually is.
Context window and model size. The draft does not mention the context window or parameter count. Given that Mythos represents a new tier above Opus, it is likely larger and likely supports a longer context, but nothing is confirmed.
Final release name. The two draft versions suggest Anthropic had not finalized whether to call it Mythos, Capybara, or something else entirely. The model that ships publicly could carry a different name from either draft.
Pricing. The leaked document says the model is very expensive to serve but includes no figures. Pricing above Opus would make it the most expensive Claude model by a significant margin.
Public release date. Considering Anthropic might go public with an IPO around October 2026, the official release of Mythos could be around that time. Bloomberg and The Information reported that Anthropic was considering an IPO as early as October 2026. However, no official connection between those two events has been confirmed.
A Reason to Be Cautious About the Hype
The leaked documents are compelling. They are also internal marketing drafts.
Every frontier AI company describes its next model as a step change, its most capable system ever, and a fundamental advance over what came before. The phrase appears in virtually every major AI announcement from every lab. OpenAI used similar language about GPT-5, which ultimately disappointed many observers when it shipped.
The pattern worth watching is whether Anthropic's characterization of Mythos as genuinely different from its existing lineup holds when independent researchers get access to benchmark the model. The cybersecurity claims in particular are extraordinary enough to warrant skepticism until they can be verified externally.
What the leak confirms is that Anthropic has completed training a new model, is testing it with early access customers, and is being cautious about release due to safety concerns. Everything beyond that is Anthropic's own description of its own work, filtered through a draft document that was never meant to be public.
What This Means for Developers and Businesses
If Mythos delivers on its described capabilities, several things change.
For developers using Claude Code, a model with dramatically higher coding scores would compound an already significant advantage over competing tools. Claude Code's enterprise market share in 2026 has been built on Opus 4.6's benchmark superiority. Mythos would extend that lead further.
For cybersecurity teams, early access to Mythos for defensive purposes could represent a meaningful capability upgrade in finding vulnerabilities before attackers do. The model's potential for proactive threat detection and vulnerability assessment is the reason Anthropic is prioritizing this access group first.
For businesses considering Anthropic's models, a new Capybara tier means more capability is coming but also higher costs. The draft is explicit that the model will be expensive. Budgeting for Mythos access will require a different calculation than current Opus pricing.
For competitors, a confirmed Anthropic model that outperforms the current state of the art across coding, reasoning, and cybersecurity increases pressure on OpenAI, Google DeepMind, and Meta to accelerate their own release timelines.
Sources
- Fortune: Exclusive first report on the leak — Original reporting by Bea Nolan
- Fortune follow-up: Cybersecurity risk angle — Expanded coverage after Anthropic confirmed the model
- The Decoder: Leaked blog post analysis — Close reading of both draft versions
- Techzine: Step-change model breakdown -- Technical detail on the Capybara tier
- Futurism: Context and skepticism — Critical perspective on the hype
- CoinDesk: Market reaction coverage -- Cybersecurity stock impact
Published: March 27, 2026. This article is based on leaked draft documents confirmed by Anthropic as real but described as early-stage drafts. All capability claims reflect Anthropic's own characterization of its model as described in those documents. Independent verification is not yet possible. For the complete April 2026 AI model landscape including GPT-5.5, Gemini 3.1 Pro, and more, see our AI Models in April 2026 guide.
Frequently Asked Questions
What is Claude Mythos?
Claude Mythos is an unreleased AI model from Anthropic, accidentally revealed through a data leak on March 26, 2026. It is also referred to internally as Capybara. Anthropic confirmed it is real and described it as a step change in capability and the most powerful model the company has built to date. It outperforms Claude Opus 4.6 on coding, academic reasoning, and cybersecurity benchmarks.
What is the Capybara tier and how is it different from Opus?
Capybara is a new model tier Anthropic is introducing above its existing Opus lineup. The leaked draft states it is larger and more intelligent than Opus models, which were until now the most powerful. This adds a fourth tier to Anthropic's current lineup of Haiku, Sonnet, and Opus. Mythos appears to be the first model within the Capybara tier.
How did the Claude Mythos leak happen?
A configuration error in Anthropic's content management system left approximately 3,000 unpublished assets in an unencrypted, publicly searchable data cache. The exposed files included draft blog posts and internal documents. Security researchers Roy Paz of LayerX Security and Alexandre Pauwels from the University of Cambridge independently discovered the exposed data. Fortune reviewed and published the findings after notifying Anthropic, which then restricted access.
Why is Anthropic not releasing Claude Mythos publicly?
Anthropic cited the model's unprecedented cybersecurity capabilities as the primary reason for a cautious release strategy. The leaked documents state the model is currently far ahead of any other AI model in cyber capabilities and could enable vulnerability exploitation at speeds that outpace defenders. The model is also described as very expensive to serve. Anthropic is currently testing it with a small group of early access customers focused on cybersecurity defense scenarios.
When will Claude Mythos be publicly available?
No confirmed public release date has been announced. Anthropic is testing it with a limited group of early access customers and described its approach as deliberate. Some analysts have speculated the release could align with Anthropic's potential IPO, which Bloomberg and The Information reported could come as early as October 2026, though no official confirmation links the two events.
What does Claude Mythos score on benchmarks?
Specific numerical scores were not included in the leaked documents. However, the draft blog post states that compared to Claude Opus 4.6, Capybara gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity. Anthropic's spokesperson confirmed meaningful advances in all three of those areas. Cybersecurity performance is described as far surpassing all existing AI models, including Anthropic's own.
Did the Claude Mythos leak affect the stock market?
Yes. Following the leak, cybersecurity stocks fell sharply. Palo Alto Networks, CrowdStrike, and Fortinet all dropped 4 to 6 percent. The iShares Expanded Tech-Software Sector ETF fell approximately 2.5 to 3 percent. The concern among investors was that a model with superior vulnerability detection capabilities could undermine existing cybersecurity products and defenses.
