Anthropic announced the availability of the Claude Fable 5.1 and Claude Mythos 5.1 models, focusing on programming, information-intensive tasks, and long-running problem-solving. Both models use the same underlying architecture, but Mythos 5.1 comes with more flexible safety restrictions and is available only to verified users and organizations in the fields of cybersecurity and life sciences.
Fable 5.1 targets general use, while Mythos 5.1 represents the version aimed at specialized businesses that may need a broader scope for testing or technical operations. Anthropic says the new model maintains the same input and output token prices as Fable 5, at $10 per million input tokens and $50 per million output tokens through the API. In contrast, the company cut the price of reading cached tokens by 75% to $0.25 per million tokens.
Lowering the Cost of Agentic Tasks
According to Anthropic, these changes give users savings of about 25% on typical workloads compared with Fable 5, with savings reaching approximately 45% on agentic tasks involving numerous tool calls and long contexts. The company says Fable 5.1 can achieve performance comparable to or higher than Fable 5 when using low or medium levels of computing capacity. By default, the model operates at a high level in Claude Code and at a medium level in Claude Cowork and Claude.ai.
Better Results in Programming and Research
Tests presented by Anthropic showed clear improvements in some areas. Fable 5.1 scored 52.6% on the Terminal-Bench-Science 0.1 test, compared with 24.7% for Fable 5. On the Terminal-Bench 4.0 programming test, Fable 5.1 scored 55.8%, while Mythos 5.1 reached 60.9%, compared with approximately 42% for Fable 5.
The company provides examples of long-running tasks, including Fable 5.1 discovering a rare software bug in the systems of investment firm Millennium that appeared approximately once per million executions. The model also used radar images collected by NASA’s Magellan mission to create a high-resolution elevation map covering about one-third of Venus. In computational biology, Mythos 5.1 helped optimize graphics processing units for seven open-source deep-learning models, accelerating some models by up to 2.5 times.
New Restrictions to Counter Model Distillation
Fable 5.1 includes mechanisms to combat model distillation, a process of collecting the outputs of an advanced model on a large scale to train another model. Anthropic says these operations may rely on thousands of fake accounts to transfer the capabilities of advanced models to systems that do not apply the same safety controls.
Under the change, new Claude API accounts will not allow users to manually modify the previous conversation context while retaining the model’s previous reasoning records in a multi-turn conversation. Anthropic says this measure closes one publicly documented method of model distillation. The company had previously announced the discovery of operations involving approximately 24,000 fake accounts and more than 1.6 million interactions with Claude, which it linked to DeepSeek, Moonshot AI, and MiniMax. The new restriction currently applies only to new accounts, with the possibility of expanding it to all accounts in future releases.
What Changes in Practice?
The most significant value in the announcement is not limited to the names of the two models, but combines lower costs for agentic use with tighter control over how previous context is accessed. This may benefit developers managing many long-running tasks with numerous calls, but it also adds a restriction to methods for testing or managing conversations through the API. The performance results cited are figures provided by Anthropic itself, so the material does not clarify how comparable they are outside the mentioned tests.
Fable 5.1 is available through the Claude platform and the services of Amazon Web Services, Google Cloud, and Microsoft Azure, as well as through the Claude API under the identifier claude-fable-5-1. Mythos 5.1, meanwhile, is available only to cybersecurity and life-sciences users who pass the Cyber Verification Program or Life Sciences Verification Program. The life-sciences program was developed in cooperation with the U.S. government, while the Claude Security product, dedicated to scanning codebases for vulnerabilities, is now powered by Mythos 5.1.