One Claude model, two levels of safeguards

Anthropic improves coding, scientific research, and long-running tasks while lowering costs with Claude Fable 5.1 and Claude Mythos 5.1.

Two names now refer to the same underlying system. Claude Fable 5.1 and Claude Mythos 5.1 share the same core capabilities but do not operate under the same safeguards. The first is available to the public and businesses, while the second reserves certain cybersecurity and life sciences applications for vetted organizations.

The distinction does not reflect a difference in model size, training, or intelligence. Fable 5.1 applies additional restrictions and may redirect certain sensitive requests to other models. Mythos 5.1 gives approved professionals greater latitude without removing the remaining safety measures.

Fable 5.1 primarily targets software development, document-based research, professional workflows, and long-running tasks involving several tools. Its effort level defaults to High in Claude Code and Medium in Claude Cowork and on Claude.ai. Users can lower this setting to prioritize speed and cost when a task does not require the deepest reasoning available.

Anthropic’s evaluations show particularly strong gains in agentic scientific research. Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, compared with 24.7% for Fable 5, 29% for Opus 5, and 22.4% for GPT-5.6 Sol under the company’s testing configuration.

On Terminal-Bench 4.0, which measures autonomous software development inside a terminal, Fable 5.1 reaches 55.8%, up from 42% for Fable 5. Mythos 5.1 reaches 60.9%. Anthropic attributes part of this difference to Fable’s safeguards, which intervene in some cybersecurity-related tasks.

Results also improve across professional workflows. Fable 5.1 scores 31.4% on AutomationBench, compared with 17.1% for its predecessor, and reaches 73.4% on CursorBench 3.2.0. It also outperforms Fable 5 and Opus 5 in the published configurations for GDPval-AA, OSWorld 2.0, and Humanity’s Last Exam. The complete figures and testing conditions appear in the official introduction to both models.

These results do not establish a universal ranking. Anthropic ran several of the evaluations using its own configurations. Some tasks trigger a model switch or receive a score of zero when safeguards intervene. The version of OSWorld used for the release was also updated, preventing direct comparisons with scores previously published for some competing models.

Early-access partners primarily emphasize the model’s ability to continue working for several hours, verify its own results, and identify the root cause of a problem. Millennium says Fable 5.1 traced a rare system crash that had remained unexplained for several years. MongoDB describes the autonomous development of a complex prototype over three days. These accounts come from companies given early access and do not replace independent evaluations.

The announcement extends beyond software development. Anthropic presents several experiments in which Fable 5.1 or Mythos 5.1 directly participated in scientific research.

For molecular design, Mythos 5.1 used open-source protein design and folding tools to propose binders for 12 targets. Two external organizations then tested the resulting designs in laboratories. Anthropic reports a hit rate approaching 50%, compared with a typical range of 10% to 15% for this type of work. For three targets, some measured binding affinities were ten times higher than the best comparable submissions to Adaptyv Bio competitions.

The results remain part of a protocol organized by Anthropic. They concern the initial design of binding molecules rather than validated medicines. High binding affinity alone does not demonstrate the effectiveness, safety, or clinical viability of a treatment.

Fable 5.1 also created a new elevation map covering about one-third of Venus using radar observations from NASA’s Magellan mission. According to Anthropic, the map distinguishes features at a scale of two to three kilometers, compared with 10 to 20 kilometers for the previous reference, with elevation accuracy improving by as much as 25%. The result has been released on Zenodo under a Creative Commons license for examination and potential use in preparations for the VERITAS and EnVision missions.

In another experiment, Mythos 5.1 modified the code of seven open-source models used in genomics and computational biology. Anthropic reports speed increases of up to 2.5 times while producing identical outputs. The company estimates that these changes could reduce computing costs by 30% to 60% for some genome-wide analyses. Anthropic plans to release the optimizations, but they were not publicly available at launch.

The expanded scientific and technical capabilities explain the creation of Mythos 5.1. Access depends on two programs. The Cyber Verification Program is intended to allow vetted professionals to perform certain defensive work that would normally be blocked. The Life Sciences Verification Program provides selected researchers with access to advanced biology capabilities through a framework developed with the US government.

Access to Mythos 5.1 is currently limited to a group of US organizations. The model also powers Claude Security, Anthropic’s service for examining codebases, identifying vulnerabilities, and suggesting patches for human review.

Fable 5.1 can now identify vulnerabilities in source code. Requests involving exploit development, penetration testing, or offensive analysis of binary files remain blocked or redirected. Anthropic reports about 60% fewer unnecessary interventions during cybersecurity sessions compared with the original safeguards applied to Fable 5.

A similar change applies to biology. Basic biology and medical questions are said to trigger 85% fewer redirects than they did when Fable 5 launched. Advanced life sciences research and development remains restricted to Mythos or handled by selected Opus models.

The 212-page system card also describes continuing limitations. Mythos 5.1 is Anthropic’s most capable released model for cybersecurity tasks but remains classified within the first risk tier of its Frontier