Anthropic introduces Claude Fable 5.1 and restricted Mythos 5.1 for advanced work

Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, presenting the models as a major upgrade for coding, knowledge work, long-running agentic tasks and scientific research. The two versions are the same underlying model but ship with different safeguards. Fable 5.1 is generally available, while Mythos 5.1 is limited to trusted access programs, with protections designed for advanced cybersecurity and life-sciences work.

The product announcement combines performance, price and safety changes. Anthropic said Fable 5.1 will cost an estimated 25% less than Fable 5 for typical token-billed workloads because of lower cache-read pricing. For highly agentic work, the company said savings can often reach about 45%. That pricing shift matters because frontier models are increasingly used for long-running tasks that read large repositories, documents and tool outputs over many turns. Lower cache costs can make the difference between using a frontier model only for a final pass and using it as the main system that carries work from research to implementation.

Anthropic also framed Fable 5.1 as a step forward in performance. The company reported stronger results than Fable 5 across benchmarks such as Terminal-Bench-Science, Terminal-Bench, Humanity’s Last Exam, OSWorld, AutomationBench and CursorBench. It also cited early-access feedback from companies including Jane Street, Cognition, Millennium, MongoDB, Red Hat, Rakuten, Block, Ramp, Canva, Hebbia and Plaid. The examples emphasize practical agent behavior: finding the root cause of rare software crashes, building prototypes over multi-day runs, reviewing complex code changes and identifying new research directions. Anthropic’s point is not simply that the model scores higher, but that it can sustain longer work with better reasoning, communication and follow-through.

The scientific portion of the release is especially notable. Anthropic said the model contributed to research tasks such as protein binder design, computational analysis and performance optimization for biological models. In one example, Claude Fable 5.1 trained a neural network to create a higher-resolution elevation map of about one third of Venus from radar imagery and prior mapping work. Anthropic said the new map improves detail from 10 to 20 kilometers down to two to three kilometers and estimates height up to 25% more accurately. In computational biology, the company said Mythos 5.1 sped up seven open-source deep-learning models by up to 2.5 times through custom GPU kernels and caching, cutting estimated GPU costs for some genome-wide analyses by 30% to 60%.

Those capabilities come with tighter controls. Anthropic said Mythos 5.1 has stronger cyber capabilities than any model it has released, though it remains in a lower risk category under the company’s Frontier Compliance Framework. The company also described testing for chemical and biological risks, prompt injection resistance and alignment behavior. Mythos 5.1 will keep safeguards that restrict access to research biology capabilities, and Anthropic said an access program for advanced biology work was developed with the U.S. government. For Fable 5.1, cybersecurity safeguards are becoming more precise, allowing the model to identify software vulnerabilities for defensive work while redirecting riskier dual-use tasks such as exploit generation and penetration testing to Opus models.

The launch is closely tied to Enterprise Frontier Safeguards, Anthropic’s new approach to giving enterprise customers both stronger monitoring and zero-data-retention-like privacy. Customers eligible for those safeguards can use Fable 5.1 and Fable 5 with zero data retention until EFS rolls out later this fall. Taken together, the release shows how frontier model companies are now packaging capability, cost and governance as one product decision. Fable 5.1 is meant to make advanced agentic work cheaper and more useful; Mythos 5.1 shows that the most sensitive capabilities will increasingly be distributed through controlled access rather than broad release.