What Is Claude Capybara? Anthropic’s New AI Tier Explained
Claude Capybara is a new model tier from Anthropic that sits above Opus in the Claude hierarchy, making it the most powerful class of AI models the company has ever built. Claude Capybara was revealed through an accidental data leak on March 26, 2026, when security researchers found approximately 3,000 unpublished files in an unsecured Anthropic data store.
The first model in the Capybara tier is called Claude Mythos, and it demonstrates dramatically higher scores in coding, reasoning, and cybersecurity compared to Claude Opus 4.6. Anthropic has described it as a “step change” in AI performance and “the most capable we’ve built to date.”
Capybara vs Mythos: Understanding the Names
The two names that surfaced in the leak serve different purposes. Capybara is a tier name — it describes a class of models, just like Haiku, Sonnet, and Opus describe other classes. Claude Mythos is the specific model that belongs to the Capybara tier.
How the Tier System Works
Think of it this way: the Opus tier contains Claude Opus 4.6, and the Capybara tier contains Claude Mythos. Future models could also be released under the Capybara tier, just as Anthropic has released multiple versions under Opus.
The naming shift from literary and musical terms (Haiku, Sonnet, Opus) to an animal — the capybara is Earth’s largest rodent — signals a fundamentally different capability class rather than an incremental version bump.
Why Two Names Caused Confusion
The leaked draft blog posts used both names interchangeably. Some documents discussed the “Capybara tier” while others referenced “Claude Mythos” as the model. Fortune’s initial reporting clarified the distinction: Capybara = tier, Mythos = model. This is now the accepted understanding across the AI community.
The Complete Claude Model Hierarchy
Before Capybara, Anthropic offered three tiers. Now the lineup has four levels, each designed for different use cases and budgets.
| Tier | Current Model | Best For | Relative Cost |
|---|---|---|---|
| Haiku | Claude Haiku 4.5 | Fast classification, chatbots, high-volume tasks | $ |
| Sonnet | Claude Sonnet 4.6 | Coding, content creation, balanced performance | $$ |
| Opus | Claude Opus 4.6 | Complex reasoning, research, architecture | $$$ |
| Capybara | Claude Mythos | Breakthrough tasks, cybersecurity, frontier reasoning | $$$$ |
What Makes Capybara Different from Opus
Capybara is not simply a bigger Opus. Anthropic’s internal documents describe it as a qualitative leap — “like taking an elevator versus climbing stairs.” The leaked benchmarks show improvements across every measured dimension, with cybersecurity capabilities being the most dramatic outlier.
Opus 4.6 already holds strong positions in industry benchmarks: 65.4% on Terminal-Bench 2.0, surpassing GPT-5.2-Codex. Capybara scores “dramatically higher” than even this baseline, though exact numbers remain under wraps.
What Claude Capybara Can Do
Anthropic’s spokesperson confirmed Capybara is a “general purpose model with meaningful advances in reasoning, coding, and cybersecurity.” The leaked documentation provides more specific details about each capability area.
Coding and Reasoning
Claude Opus 4.6 already competes with or outperforms GPT-5.4 on most coding benchmarks. Capybara takes this further with what the internal docs describe as dramatically higher coding scores. For academic reasoning, the improvement is similarly significant — suggesting architectural innovations rather than just scaling existing approaches.
The leaked materials also mention the ability to “create deep connective tissue between ideas and knowledge.” This points to cross-domain reasoning capabilities that go beyond answering questions in isolation, potentially enabling the model to synthesize insights across entirely different fields.
Cybersecurity Capabilities
This is the capability that generated the most headlines. Anthropic’s own internal assessment states the model is “currently far ahead of any other AI model in cyber capabilities.” The system can identify and exploit software vulnerabilities at speeds that outpace human defenders.
The dual-use nature of this capability is what makes it both groundbreaking and concerning. The same model that could help security teams find bugs before attackers do could also be used to discover and exploit zero-day vulnerabilities at unprecedented scale.
Performance Benchmarks
While Anthropic has not published official benchmark numbers for Capybara, the leaked documents claim “dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity” compared to Opus 4.6. For context, here is where Opus 4.6 currently stands against competitors:
| Benchmark | Claude Opus 4.6 | GPT-5.4 | Leader |
|---|---|---|---|
| SWE-Bench | 80.8% | 77.2% | Opus |
| SWE-Bench Pro | ~45.9% | 57.7% | GPT-5.4 |
| Terminal-Bench 2.0 | 65.4% | — | Opus |
| GPQA Diamond | Leading (+3.5 pts) | — | Opus |
| OSWorld | 72.7% | 75% | GPT-5.4 |
Capybara is expected to surpass Opus across all of these benchmarks, though independent verification is still pending.
How Claude Capybara Was Discovered
The Capybara tier was never meant to be public knowledge — at least not yet. Its existence came to light through one of the more ironic security incidents in recent AI history.
The Data Leak
Security researchers Roy Paz from LayerX and Alexandre Pauwels from Cambridge discovered approximately 3,000 unpublished assets in an unsecured data store connected to Anthropic’s content management system. The files included draft blog posts, internal documentation, and details about the Capybara tier and Claude Mythos model.
Fortune broke the story on March 26, 2026, after notifying Anthropic about the exposed files. The company restricted access to the data store following Fortune’s inquiry.
Anthropic’s Response
An Anthropic spokesperson attributed the leak to “human error” in the configuration of their content management system. The company confirmed the model’s existence and that it was in testing, but emphasized it was “expensive to run and not yet ready for general release.”
The irony was not lost on commentators: an AI safety-focused company accidentally leaking its most powerful model through a basic security misconfiguration. The incident also exposed details about a planned invite-only CEO summit where Anthropic’s Dario Amodei would meet European business leaders.
Cybersecurity Impact and Stock Market Reaction
The leak had immediate real-world consequences beyond the AI industry. The revelation that an AI model could outperform human cybersecurity experts sent shockwaves through financial markets.
Why Cybersecurity Stocks Dropped
Following the leak, major cybersecurity companies saw significant stock declines. CrowdStrike (CRWD) fell approximately 7%, Palo Alto Networks (PANW) dropped about 6%, and Fortinet (FTNT) lost 4-6%. The IGV cybersecurity ETF also declined.
Investors reacted to the implication that AI models capable of finding and exploiting vulnerabilities faster than defenders could fundamentally disrupt the cybersecurity industry’s business model.
The Dual-Use Problem
Anthropic acknowledges the tension directly. The same capabilities that make Capybara valuable for defensive security also make it dangerous in the wrong hands. Their planned solution is a phased rollout starting with cyber defense organizations, giving defenders a “head start in improving the robustness of their codebases.”
This approach mirrors industry precedent: OpenAI’s GPT-5.3-Codex was also classified as having “high capability” for cybersecurity tasks, and the company took similar precautions with its release.
Claude Capybara vs GPT-5 and Competitors
The AI landscape in 2026 is defined by rapid capability jumps. Capybara enters a competitive field, but the leaked benchmarks suggest it could establish a significant lead.
Capybara vs GPT-5
OpenAI’s GPT-5, released in August 2025, was widely considered a disappointment relative to the company’s promises. Claude Opus 4.6 already outperforms GPT-5 on several key benchmarks. With Capybara described as “dramatically” better than Opus 4.6, the gap between Anthropic and OpenAI could widen substantially.
GPT-5.4, the latest iteration, has caught up in some areas — particularly coding (SWE-Bench Pro) and computer use (OSWorld). But if Capybara’s leaked performance claims hold up, it would represent a generational leap over all current competitors.
The Broader AI Race
Claude Code and Claude Cowork have already established Anthropic as a leader in developer tooling. Capybara could cement the company’s technical lead across the board. However, caution is warranted — curated benchmark results from leaked internal documents don’t always translate to real-world performance at scale.
When Will Claude Capybara Be Available
Capybara is not available to the general public yet, and Anthropic has been deliberately vague about timelines.
Current Access Status
The model is in restricted early testing with a “small group of early access customers.” Anthropic has prioritized cybersecurity defense organizations for initial access, consistent with their stated goal of giving defenders a head start.
Expected Release Timeline
Based on available information and Anthropic’s track record:
- Q2 2026: Expanded testing with additional enterprise customers
- Q3 2026: Potential limited beta or waitlist opening
- October 2026: Most likely window for general availability
The October 2026 timing aligns with Anthropic’s potential IPO, reported by Bloomberg at a valuation of $60 billion or more. A major model launch would provide favorable momentum for a public offering.
API Access and Pricing
When Capybara becomes generally available, accessing it through the Claude API will require only changing the model parameter — the SDK and request format remain identical to current Claude models. No official pricing has been announced, but community estimates range from 2-3x Opus pricing (conservative) to 4-5x (aggressive). The leaked docs confirmed the model is “expensive to run.”
Questions About Claude Capybara
Is Claude Capybara the same as Claude Mythos?
Not exactly. Capybara is the tier name, similar to how Opus and Sonnet are tier names. Claude Mythos is the first specific model within the Capybara tier. Future Capybara-tier models may have different names.
Why did Anthropic name it Capybara?
Previous tiers used literary and musical names: Haiku, Sonnet, Opus. The shift to an animal name signals a fundamentally new capability class. The capybara — Earth’s largest rodent — suggests something unexpectedly large and capable.
How much will Claude Capybara cost?
No official pricing yet. Estimates range from 2-3x to 4-5x current Opus pricing. Anthropic confirmed the model is expensive to run, so expect premium pricing when it launches.
Can Claude Capybara be used for hacking?
The model has unprecedented cybersecurity capabilities, including the ability to find and exploit software vulnerabilities. Anthropic is releasing it first to cyber defense organizations specifically to mitigate misuse risks.
When can I use Claude Capybara?
The most likely window for general availability is late 2026, potentially October, aligned with Anthropic’s expected IPO timeline. Currently it is in restricted testing with select enterprise customers.
Is Claude Capybara better than GPT-5?
Based on leaked internal documents, Capybara dramatically outperforms Claude Opus 4.6, which itself already beats GPT-5 on multiple benchmarks. If the claims hold up in independent testing, Capybara would represent a significant lead over all current competitors.
What went wrong with the Anthropic data leak?
A CMS misconfiguration left approximately 3,000 internal files publicly accessible. Security researchers Roy Paz and Alexandre Pauwels found the files, and Fortune reported the story on March 26, 2026. Anthropic attributed it to human error.