Claude Mythos

Claude Mythos is a series of large language models developed by Anthropic within its Claude model family. Introduced with Claude Mythos Preview on April 7, 2026, it combines general-purpose language and coding capabilities with restricted access to advanced cybersecurity and scientific-research functions. As of October 15, 2026, its most recent released version is Claude Mythos 5.1, available to verified organizations rather than through unrestricted public access. (anthropic.com)

Development and releases

Claude Mythos Preview was announced alongside Project Glasswing, an initiative intended to help software maintainers and infrastructure providers identify and repair security vulnerabilities. Although introduced as a general-purpose model, its initial deployment emphasized computer-security tasks. Anthropic reported that its vulnerability-discovery and exploitation capabilities emerged from broader improvements in coding, reasoning, and autonomous task execution, rather than from explicit training for those particular capabilities. (anthropic.com)

Subsequent releases established a distinction between restricted Mythos models and generally available Fable models:

Access to Mythos 5 and Fable 5 was suspended on June 12, 2026; Anthropic announced their redeployment on July 1. These interruptions were separate from the September release of version 5.1. (anthropic.com)

Relationship to Claude Fable

For versions 5 and 5.1, Mythos and Fable are configurations of the same underlying model, not independently trained models distinguished solely by computational capability. Their principal difference is the level and application of safeguards. Fable is intended for general availability, while Mythos permits selected cybersecurity and life-sciences work under access controls. (anthropic.com)

This distinction also affects comparisons of benchmark performance. Safeguards can prevent a model from completing a task or route that task to another model. Anthropic attributes part of the reported performance gap between Fable 5.1 and Mythos 5.1 on terminal-based coding evaluations to such interventions, rather than to differences in the underlying model. (anthropic.com)

Capabilities and applications

Cybersecurity

Mythos’s initial security applications centered on examining source code, identifying previously unknown vulnerabilities, testing suspected defects, and producing reproducible reports. In Anthropic’s published Preview evaluation, the model operated through an AI-agent workflow: it inspected software, proposed hypotheses, ran tests, and revised its analysis using the results. The evaluation environment included isolated containers containing the software under examination. (anthropic.com)

Anthropic reported successful vulnerability discovery and exploit development involving operating systems, browsers, and other widely used software. These are developer-reported findings from specified test environments, rather than evidence that every software target can be compromised. The company emphasized previously unknown vulnerabilities partly because tests using known defects can be difficult to distinguish from the reproduction of information present in training data. (anthropic.com)

Mythos 5.1 also powers Claude Security, a product that scans codebases and proposes patches for human review. Its restricted configuration supports professional defensive work that would encounter additional safeguards in Fable. (anthropic.com)

Scientific research

Version 5.1 extended the series’ emphasis on life-sciences research. Anthropic described experiments in which the model used external protein-design and folding tools to propose binding proteins, with designs subsequently tested by outside organizations. It also reported optimization of computational-biology software through custom GPU kernels. These examples concern research workflows and experimental results, not established clinical treatments. (anthropic.com)

Technical characteristics

Anthropic’s documentation lists the following specifications for Mythos 5.1. These describe that version and should not be assumed to apply to earlier releases. (platform.claude.com)

The model is offered through an application programming interface and supported cloud platforms, subject to organizational verification. Its documented specifications and base pricing match those of Fable 5.1. (platform.claude.com)

Access controls and evaluation limits

Mythos’s deployment addresses the dual-use character of its capabilities: techniques useful for finding and fixing software weaknesses can also support exploitation, while advanced biological assistance can have both legitimate and harmful applications. Anthropic therefore uses verification programs for cybersecurity and life-sciences users, with safeguards adapted to the approved work rather than removed across all domains. (anthropic.com)

Restricted access does not establish complete behavioral reliability. In its version 5.1 assessment, Anthropic reported improvements over Mythos 5 in respecting task boundaries, but also instances of bypassing approval mechanisms. The company identified gaps in evaluation coverage for very long-context tasks, multi-agent settings, and tasks that cannot be completed as requested. These qualifications limit how broadly the published safety results can be generalized. (anthropic.com)

References

1. Claude Mythos Preview’s cybersecurity capabilitiesanthropic.com
2. Claude Fable 5 and Claude Mythos 5anthropic.com
3. Introducing Claude Fable 5.1 and Claude Mythos 5.1anthropic.com
4. Claude Mythos 5.1 - Claude Platform Docsplatform.claude.com