What Is Claude Mythos Preview?
A comprehensive guide to Anthropic's most powerful AI model — last updated April 24, 2026.
Claude Mythos Preview is an unreleased frontier AI model developed by Anthropic, officially announced on April 7, 2026. It represents a "step-change" in AI capability, particularly in autonomous cybersecurity tasks including identifying and exploiting zero-day vulnerabilities. Anthropic decided not to release it publicly due to its advanced offensive capabilities. Access is limited to 12 founding partners through Project Glasswing, a controlled initiative for defensive cybersecurity.
What Is Claude Mythos Preview?
Claude Mythos Preview is the internal designation for a frontier AI model developed by Anthropic that sits in a new tier above the Claude Opus product line. According to leaked internal documents that surfaced in late March 2026, Anthropic uses the codename "Capybara" to refer to this capability tier, positioning it as a generational leap rather than an incremental improvement over existing models.
The model was officially announced on April 7, 2026, alongside the launch of Project Glasswing. Anthropic described it as having "step-change" performance improvements across software engineering, mathematical reasoning, and — most notably — autonomous cybersecurity capabilities.
Unlike Claude Opus 4.6 and the subsequently released Claude Opus 4.7, Mythos Preview is not available through the Claude API, the Claude web interface, or any public-facing product. Anthropic made the deliberate decision to restrict access specifically because of the model's cybersecurity capabilities.
Who Made It?
Claude Mythos Preview was developed by Anthropic PBC, the AI safety company founded in 2021 by former OpenAI researchers, including Dario Amodei (CEO) and Daniela Amodei (President). Anthropic is the creator of the Claude family of AI models and is known for its focus on AI safety research, including its Constitutional AI approach.
Anthropic has raised over $10 billion in funding and counts Amazon and Google among its major investors and cloud infrastructure partners. The company is headquartered in San Francisco.
What Can It Do?
Based on Anthropic's published system card, Claude Mythos Preview demonstrates exceptional capability across several domains:
Software Engineering
The model achieves 93.9% on SWE-bench Verified, the industry-standard benchmark for autonomous software engineering. This represents a significant improvement over Claude Opus 4.6 (80.8%) and surpasses reported scores for GPT-5 (~85%). See our full benchmark comparison for detailed data.
Cybersecurity
This is where Mythos Preview is most distinctive. The model achieves 100% on Cybench (pass@1), a capture-the-flag cybersecurity benchmark, and 0.83 on CyberGym, which measures both offensive and defensive capabilities. During internal testing, the model:
- Discovered thousands of critical vulnerabilities across major operating systems and web browsers
- Found a 27-year-old vulnerability in OpenBSD that had gone undetected by human security researchers
- Developed functional proof-of-concept exploits from vulnerability discoveries in the Firefox 147 JavaScript engine
- Demonstrated the ability to autonomously move from vulnerability identification to exploit development with minimal human guidance
Mathematical Reasoning
Mythos Preview scores 97.6% on USAMO 2026 (USA Mathematical Olympiad), compared to 42.3% for Claude Opus 4.6 — a more than twofold improvement that observers describe as unprecedented in the history of AI benchmarking.
General Autonomy
On OSWorld, which measures a model's ability to autonomously navigate computer desktops and complete tasks, Mythos Preview achieves 79.6% — a meaningful improvement over previous models, though less dramatic than its cybersecurity gains.
Why Isn't It Public?
Anthropic made the explicit decision not to release Mythos Preview to the public, citing its extraordinary offensive cybersecurity capabilities. According to the system card, the model's ability to autonomously identify vulnerabilities, develop exploits, and chain them together means that widespread access could pose significant risks to critical software infrastructure.
The company determined that the potential for defensive benefit — finding and patching vulnerabilities before attackers do — was best realized through controlled access to trusted organizations, rather than through broad public release. This led to the creation of Project Glasswing.
Instead of Mythos Preview, Anthropic released Claude Opus 4.7 on April 16, 2026, as its publicly available frontier model. Opus 4.7 features improvements in software engineering and reasoning but does not include the specialized cybersecurity capabilities of Mythos Preview.
Project Glasswing
Project Glasswing is the controlled-access initiative through which Anthropic provides Claude Mythos Preview to selected organizations. The program has 12 founding partners:
- Amazon Web Services (AWS)
- Apple
- Broadcom
- Cisco
- CrowdStrike
- JPMorgan Chase
- The Linux Foundation
- Microsoft
- NVIDIA
- Palo Alto Networks
- Anthropic (internal)
Beyond these 12 founding partners, Anthropic reports that over 40 additional organizations responsible for critical software infrastructure have received access. The full list of additional participants has not been disclosed. See our complete partner analysis.
Benchmark Performance
For a detailed comparison of Claude Mythos Preview against other frontier models, see our full benchmarks page. The key takeaway: Mythos Preview represents what independent observers describe as a "step-change" rather than incremental improvement, particularly in cybersecurity and mathematical reasoning.
Timeline of Events
The story of Claude Mythos Preview is still unfolding. For a complete chronological record of key events — from the March 2026 Capybara leak to the latest developments around unauthorized access reports — see our full timeline.
Government Involvement
The model's cybersecurity capabilities have drawn attention from the U.S. government. According to reports, the NSA is actively using Mythos Preview for cyber defense operations, even as the Pentagon has formally designated Anthropic as a "supply-chain risk" following a contract dispute.
White House engagement has included discussions between Anthropic leadership and the Chief of Staff and Treasury Secretary regarding safe use of the model in government agencies. The Office of Management and Budget (OMB) has reportedly been working to establish safeguards for broader government access.
Frequently Asked Questions
Is Claude Mythos Preview available to the public?
No. As of April 2026, Claude Mythos Preview is not publicly available. Anthropic has stated it has no current plans to make the model generally accessible due to its advanced cybersecurity capabilities and the potential for misuse.
What is Project Glasswing?
Project Glasswing is Anthropic's controlled-access program that provides Claude Mythos Preview to 12 founding partners and over 40 additional organizations. Access is restricted to defensive cybersecurity purposes, specifically identifying and patching vulnerabilities in critical software infrastructure.
How does Mythos Preview compare to GPT-5?
Based on published benchmarks, Claude Mythos Preview outperforms GPT-5 on SWE-bench Verified (93.9% vs ~85%), Cybench CTF (100% vs ~70%), and USAMO 2026 (97.6% vs ~55%). However, direct comparisons should account for different evaluation conditions.
What is the Capybara tier?
Leaked internal documents revealed that Anthropic uses "Capybara" as a tier designation above the existing Claude Opus line. Mythos Preview is the first model in this tier, representing a step-change in capability rather than incremental improvement.
Can I get API access to Mythos Preview?
No. There is no public API access to Claude Mythos Preview. Even API customers with existing Anthropic contracts cannot access this model. Access is exclusively through Project Glasswing partnerships.
What vulnerabilities has Mythos found?
According to Anthropic's system card, Mythos Preview has discovered thousands of critical vulnerabilities across major operating systems and web browsers during testing. Notable discoveries include a 27-year-old vulnerability in OpenBSD and exploitable flaws in the Firefox 147 JavaScript engine.
Is Mythos Preview safe?
Anthropic describes the model as having extraordinary capabilities that require careful controls. The decision not to release it publicly was made specifically because its offensive cybersecurity capabilities could outpace current defenses if broadly available. Project Glasswing includes safety protocols and monitoring.
When will Mythos be publicly released?
Anthropic has not announced any public release date. The company has stated it does not have current plans to make "Mythos-class" capabilities generally available. The publicly available Claude Opus 4.7, released April 16, 2026, offers improved capabilities without the specialized cybersecurity features.