Anthropic has released Claude Fable 5.1 as its latest generally available model for coding, knowledge work and long-running AI agents, while introducing Claude Mythos 5.1 for vetted organizations working in areas such as cybersecurity and life sciences. Both versions are built on the same underlying model, but operate with different safeguards depending on the use case.
The most commercially significant change may not be a benchmark score. Fable 5.1 cuts the price of reading cached context by 75%, to $0.25 per million tokens. Anthropic estimates that this can reduce the cost of typical workloads by about 25% and highly agentic workloads by as much as 45%.
That matters because the economics of AI are changing. As companies move from asking models individual questions to allowing them to work continuously across applications, codebases and business processes, the cost of maintaining and repeatedly accessing context becomes increasingly important.
Built for Work That Takes Hours, Not Seconds
Anthropic is positioning Fable 5.1 around a different definition of AI performance.
Instead of optimizing primarily for isolated prompts, the model is designed for tasks that can extend across hours, multiple applications and several stages of execution. It can plan work, use tools, recover from failures and continue operating with less human supervision.
This includes autonomous coding sessions, browser-based tasks, enterprise research and complex knowledge work.
The distinction is important because the next phase of enterprise AI may not be defined by how intelligently a model answers a question, but by how reliably it can complete an entire job.
Anthropic is effectively moving the product from an AI assistant toward an AI worker.
The Economics of Persistent Agents
This is where the 75% reduction in cache-read pricing becomes particularly important.
AI agents frequently need to maintain large amounts of context: project documentation, previous actions, codebases, instructions and intermediate results. If that information has to be repeatedly processed at full cost, autonomous systems become expensive to operate at scale.
Caching allows frequently reused context to be accessed more cheaply.
With Fable 5.1, Anthropic is making that architecture significantly more economical. The company says the reduction can lower total costs by approximately 25% for typical workloads and up to roughly 45% for highly agentic workloads.
That creates a potentially important shift in the economics of autonomous software.
A model that is slightly more capable but dramatically cheaper to operate across thousands of long-running tasks can be more valuable to an enterprise than a model that wins a benchmark but remains expensive to deploy continuously.
Performance Is Moving Beyond Coding
Anthropic reports improvements across coding, automation and knowledge work.
In its reported evaluations, Fable 5.1 scored 55.8% on Terminal-Bench 4.0, compared with 42.0% for Fable 5. On AutomationBench, it reached 31.4%, compared with 17.1% for its predecessor. Its reported GDPval-AA v2 score for knowledge work reached 1,853, compared with 1,723 for Fable 5.
These figures are Anthropic-reported benchmarks, so they should be interpreted in that context. The company also notes that production safeguards affected some evaluations and that certain OSWorld results are not directly comparable with earlier releases.
The more interesting signal is therefore not one individual benchmark.
It is the combination of capability, persistence and lower operating cost.
That combination is what could make advanced models increasingly practical for real business processes.
Mythos 5.1 Shows Where the Ceiling Is
Anthropic is also making Mythos 5.1 available through restricted access programs for vetted organizations.
The underlying model is the same as Fable 5.1, but Mythos operates with more permissive safeguards for specific high-risk research environments. Anthropic says the program is intended for organizations working in areas including advanced cybersecurity and biology.
Anthropic reports that Mythos 5.1 reaches 60.9% on Terminal-Bench 4.0 under its more permissive cybersecurity configuration.
The company is also highlighting research capabilities beyond software. It says Mythos 5.1 has been used to design experimentally validated protein binders, while Fable 5.1 trained a neural network that generated a higher-resolution elevation map covering roughly one-third of Venus.
These examples point toward a broader ambition: frontier AI is increasingly being positioned not only as a tool for generating text and code, but as a system capable of contributing to scientific and technical research.
Anthropic Is Also Changing Where Monitoring Data Lives
Alongside the model launch, Anthropic is introducing Enterprise Frontier Safeguards, or EFS.
The architecture is designed for organizations that need stronger control over sensitive AI monitoring data. Eligible customers can store misuse-detection and monitoring data within their own cloud infrastructure, using customer-managed encryption keys, access controls and audit logging.
Anthropic says automated systems can analyze the data for serious misuse, while alerts are sent to the customer for review. Human review is, by default, handled by the customer rather than Anthropic.
This is particularly relevant to highly regulated organizations.
Fable-class models have advanced capabilities that require safety monitoring, but the same organizations deploying them may be unwilling or unable to place sensitive conversational data under a provider-controlled retention architecture.
EFS attempts to solve that tension by separating frontier-model capability from centralized control of monitoring data.
A Different Enterprise AI Architecture
The significance of Fable 5.1 therefore extends beyond the model itself.
Anthropic is simultaneously addressing three barriers to enterprise adoption:
Capability: the model can handle increasingly complex and long-running tasks.
Economics: cheaper cache reads reduce the cost of persistent agent workloads.
Governance: organizations gain greater control over the infrastructure containing sensitive monitoring data.
That combination is strategically important.
The enterprise AI market is moving away from simply asking which model is smartest. Companies increasingly need to know which model can perform useful work reliably, at an acceptable cost and under governance structures that their legal and security teams can approve.
Fable 5.1 is designed around that equation.
The Real Competition Is Agent Economics
The AI industry’s next competitive phase may therefore be less about headline intelligence and more about cost per completed task.
An agent that spends hours working through a software project, analyzing documents, operating tools or performing research creates a fundamentally different cost structure from a chatbot responding to individual prompts.
Every improvement in context efficiency, inference cost, reliability and autonomy compounds across those workloads.
Anthropic’s 75% reduction in cache-read pricing is significant precisely because it targets one of those underlying costs.
The company is not simply making Claude smarter.
It is trying to make long-running AI work economically viable.
And that may ultimately matter more than another benchmark victory.
