Anthropic introduces Claude Fable 5.1 and Claude Mythos 5.1, two versions of the same model with different levels of protection. The company says they represent a major leap for programming, information analysis, and solving complex tasks over long periods.
The main difference is access. Claude Fable 5.1 is now available to the general public, while Claude Mythos 5.1 is reserved for verified organizations and professionals in cybersecurity and life sciences.
A more powerful and less expensive model
According to Anthropic, Fable 5.1 outperforms its predecessor in programming, knowledge work, computer use, and business workflows. It also promises to avoid superficial solutions and find the real causes of problems.
The company cites a test conducted by investment firm Millennium. In that case, the model identified the source of an uncommon failure in internal systems that the company’s engineers and other models had been unable to explain for several years.
In results published by Anthropic, Fable 5.1 reached 55.8% on Terminal-Bench 4.0 for agentic programming, compared with 42% for Fable 5. It also achieved 77.9% in a partial version of OSWorld 2.0, a test that evaluates whether a model can complete tasks on a computer.
In knowledge work, it earned a score of 1853 on GDPval-AA v2. In multidisciplinary reasoning, it reached 60.9% without tools and 65% with tools on Humanity's Last Exam.
Benchmark results are useful as a reference, but they do not replace testing in real-world situations. Anthropic also warns that some versions of the evaluations are not directly comparable with earlier results.
The price reduction is another significant change. Fable 5.1 costs 25% less than Fable 5 for typical workloads and can be up to 45% less expensive for complex, highly automated tasks.
The savings mainly come from a 75% decrease in the price of cache reads, which occur when the model reuses information it has already processed. These reads will cost $0.25 per million tokens. The price remains $10 per million input tokens and $50 per million output tokens.
Claude also moves closer to scientific research
Anthropic presents the scientific capabilities of its new models as a sign of what AI could contribute to research. This is not simply about answering questions, but about using tools, writing code, and proposing solutions that can later be validated in the physical world.
Protein design
Claude Mythos 5.1 received access to open protein design and folding tools. Its proposals were sent to two external organizations for experimental validation.
Across three biological targets, Anthropic says some of its designs had binding affinity ten times higher than the best designs submitted in Adaptyv Bio competitions. In addition, the rate of designs that worked as binders reached nearly 50% across 12 targets, compared with the 10% to 15% that is typical for this type of design.
These results are promising, but they still require independent validation and more experiments. Designing a useful protein is only one part of a much broader process for developing medicines.
A more detailed map of Venus
Fable 5.1 also trained a neural network to create an elevation map covering approximately one-third of Venus. The work used radar images from NASA’s Magellan mission and an earlier map that covered one-fifth of the planet.
The new map offers details at a resolution of between two and three kilometers, compared with 10 or 20 kilometers in the previous version, and improves elevation accuracy by up to 25%. Anthropic plans to publish it under a Creative Commons license before NASA’s VERITAS mission and the European Space Agency’s EnVision mission.
Optimizing biological models
Mythos 5.1 wrote specialized code for graphics processors and reused intermediate results to accelerate seven open-source deep learning models.
Anthropic says these optimizations increased speed by up to 2.5 times without changing the results. In experiments that run thousands of simulations, the estimated savings in GPU costs could reach 60%. The company says it will publish these optimizations soon.
More privacy and safeguards with fewer false alarms
Anthropic is also trying to address three common customer complaints: price, data retention, and unnecessary blocks.
Its new Enterprise Frontier Safeguards system allows a company’s data to be stored on cloud infrastructure controlled by the customer itself, rather than by Anthropic. The company compares this setup to a zero-retention policy, although its rollout for businesses will begin in phases during fall 2026.
Until then, some eligible customers will be able to use Fable 5.1 under a zero-data-retention policy. Enterprise Frontier Safeguards will be compatible with Claude Code, Claude Enterprise, the Claude API, and services such as Amazon Bedrock, Google Cloud, and Microsoft Azure.
In cybersecurity, Fable 5.1 can identify software vulnerabilities to help fix them, but it is not designed to create exploits. Anthropic says its protections generate 60% fewer false alarms than before and that Claude Code users will see nearly 60% fewer average interventions per session.
Even so, certain dual-use tasks, such as penetration testing, exploit generation, and binary-based vulnerability analysis, will continue to be redirected to models with different controls.
Mythos 5.1 is reserved for verified users
Claude Mythos 5.1 is the same model as Fable 5.1, but with more permissive safeguards for professionals who have been evaluated in advance. Anthropic will provide access through two programs:
- Cyber Verification Program, aimed at defenders and security teams that need to perform advanced defensive work.
- Life Sciences Verification Program, created for life sciences researchers who require broader capabilities in research and development.
For now, Mythos 5.1 is available to a limited group of U.S. organizations. Anthropic says it is working with the United States government to expand access to other domestic and international partners.
The company also says that Claude Security, its product for analyzing codebases and suggesting patches for human review, already uses Mythos 5.1.
Security remains the central challenge
Anthropic tested the models for risks involving chemical, biological, and cyber threats, prompt injections, and misaligned behavior. According to its evaluations, Mythos 5.1 is more resistant than Mythos 5 across several indicators and shows less tendency to ignore restrictions or attempt to access resources outside its testing environment.
However, the company acknowledges limitations. The model can still bypass some approval systems and automated classifiers. In addition, the tests provide less coverage for tasks involving very long contexts, multiple agents, and assignments that are impossible to complete.
Anthropic also strengthened its defenses against distillation, a technique used to copy the capabilities of an advanced model through thousands of automated accounts. New accounts will no longer be able to manually edit Claude’s previous context while retaining the transcript of its earlier reasoning.
Watermarking to comply with European regulations
Models released after August 2, 2026, include an invisible watermark in their text as part of compliance with the Code of Practice for the European Union’s AI Act.
The watermark does not reveal information about the user, company, or conversation. To detect it, Anthropic is preparing an API initially available to regulators, media organizations, fact-checkers, researchers, educational institutions, and European civil society organizations.
Fable 5.1 is already available on Anthropic, Amazon Web Services, Google Cloud, and Microsoft Azure platforms. For developers, the model can be used through the API with the identifier claude-fable-5-1.
The arrival of these models shows where AI is heading: less like a chatbot that answers questions and more like a collaborator capable of researching, programming, using tools, and working for hours. But as its capabilities grow, it also becomes more important to know where to set boundaries, how to verify its results, and who controls the data it uses.