Large language model (LLM) development has been moving quickly, with new model families arriving in short cycles and older naming schemes getting replaced as capability tiers shift.
When Anthropic first released Claude Fable in June 2026, it was the public version of a Mythos-class model that had earlier been limited to a smaller group of users.
That first Fable release was meant to put a more capable coding and knowledge-work model into general use while keeping extra safeguards in place. Access was interrupted shortly after launch and then restored in early July.
A few months later, on September 1, 2026, the company introduced 'Claude Fable 5.1.'
Fable 5.1 is described as an update aimed at coding and knowledge work, with particular attention to long-running tasks that stretch across many steps.
Anthropic says the model holds up better over extended sessions and that its research-related results are an early look at how models might assist scientific work. The public model remains the safeguarded version.
Claude Mythos 5.1 uses the same underlying system with fewer restrictions in some areas and is limited to trusted access programs for cyber defenders and life scientists.
On the company’s own tests, Fable 5.1 posted 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5’s 24.7%. On Terminal-Bench 4.0 it scored 55.8%, compared with 42.0% for Fable 5.
Other reported figures include higher scores on knowledge-work and computer-use evaluations, along with gains on business workflow and coding benches.
Those numbers come from Anthropic's launch materials, so they should be read as the company’s measurements rather than independent replication.
Performance is also tied to effort settings.
At lower effort levels, Fable 5.1 is said to reach similar or better results than Fable 5 while using less compute.
Base token prices are unchanged at $10 per million input tokens and $50 per million output tokens. Cache reads, however, are 75% cheaper than they were on Fable 5.
Anthropic estimates that this cuts practical cost by about 25% on typical workloads and by as much as 45% on highly agentic ones, where the model rereads large amounts of previously processed context.
The company also reported narrower changes to how refusals work.
Cybersecurity safeguards are said to flag benign requests about 60% less often. On basic biology and medical questions, the fallback rate has been reduced by around 85%.
Fable 5.1 is available through existing Claude products and cloud platforms. Mythos 5.1 remains limited to approved programs.
Alongside the model update, Anthropic announced 'Enterprise Frontier Safeguards,' or EFS. The company says EFS is meant to give enterprise customers privacy equivalent to zero data retention, with activity data kept in the customer’s own cloud account under customer-managed keys, while still running automated checks for adversarial use.
Detection signals go to the customer rather than depending on Anthropic staff review.
Taken together, Fable 5.1 is a point release rather than a new model class. It keeps the public Mythos-class line that began with Fable 5, raises several of the company’s coding and science-agent scores, and lowers the cost of cache-heavy work.
The more open Mythos variant stays restricted, and the enterprise privacy controls are still in a phased rollout. How much of the reported gain shows up in ordinary day-to-day use will depend on independent testing and on whether the cheaper cache pricing and looser false-positive rates hold up outside Anthropic’s own benches.
