Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on September 1st. Both use the same underlying model, with the main differences lying in access scope and security policies: Fable 5.1 is aimed at general professional users and developers, but it will implement restrictions or switch to a less capable model for high-risk requests in areas such as network security, biology, and chemistry; Mythos 5.1 retains more comprehensive capabilities and is only available to a small number of approved institutions.
Fable 5.1 has been provided to users Pro, Max, Team, and Enterprise. It can also be invoked through channels such as Claude Platform as well as AWS, Google Cloud, Microsoft Foundry, etc. API is identified as claude-fable-5-1; the price remains at $10 per million inputs and $50 per million outputs. The actual cost reduction lies in cache reading: it has dropped to $0.25 per million Token, which is 75% lower than $5 per million Fable. According to Anthropic, the total cost for typical workloads can be reduced by about 25%, and for highly proxyed tasks, it can be reduced by up to about 45%, but this depends on the context and reuse ratio, and it is not a uniform discount.
Rather than viewing this update as another round of ranking competition, it is more appropriate to see it as Anthropic reevaluating "how long models should operate." Fable 5.1 is aimed at tasks that span across applications and last for hours or even days: reading large codebases, taking on backlogged work orders, following up on requests in Slack, operating browsers, and recovering after failures. Model capabilities are still important, but the cost of sustainable operation, process readability, and security degradation also begin to determine the value of a product.
The same mold is divided into two channels; the capabilities are no longer limited to just on and off.
Anthropic claims that Fable 5.1 is its strongest general-purpose model, with a focus on enhancing long-term programming, knowledge-based work, and visual understanding of documents. It is capable of planning, invoking tools, writing tests, and checking outputs on its own in multi-stage tasks. It can also understand files as well as charts, tables, and diagrams within PDF. Official customer cases highlight its performance in complex fault localization and multi-day prototype development, but these cases come from specific environments and cannot replace the enterprise's own regression testing.
What is more noteworthy is the relationship between Fable and Mythos. In the past, models in high-risk areas were simply categorized as "completely unusable" or "usable by a limited number of people." This time, Anthropic has applied two different sets of policies to the same underlying capability: ordinary Fable requests that trigger network security restrictions may be forwarded to Opus 4.8; biology-related requests may be forwarded to Opus 5. API customers will need to configure their own new Fallback API. The forwarded requests will not be charged at the Fable rate, but the response capabilities, latency, and repeatability may change.
Mythos 5.1 is used for verified network defense and life science research. The life science verification program has been launched on a invite-only basis; the network security verification program will in the future include Mythos access, which currently only covers a group of American institutions. Officials have clearly stated that the scope of access is still very limited, therefore it cannot be stated that 'all researchers have unlocked the full model.' Claude Security is already running on top of Mythos 5.1 and represents a practical product implementation, but it is also subject to restrictions on access and usage.
Anthropic claims that the new version of biosecurity protection reduces interference with legitimate requests by 85% compared to Fable's initial version. At the same time, it still prevents penetration testing, vulnerability exploitation, and binary vulnerability scanning, and downgrades the handling of dual-use biochemical issues. This is an internal result of reducing false positives, but it does not mean that the risk of missed detections has disappeared. If enterprises rely on automatic switching, they must record the actual response models in their logs; otherwise, the same evaluation may mix results from different models together.
Caching becoming cheaper will encourage longer context retention, but a 30-day retention period remains a key factor in procurement.
Proxy tasks often repeatedly read from the same codebase, institutional documents, and historical conversations. When the input unit price remains unchanged, reducing cache reads by 75% can significantly lower the cost of repeated loading. If most of the context for a task can be found in the cache, it can save nearly as much as the highly proxyed upper limit mentioned by Anthropic; however, if the materials change every day and cannot be reused, the actual reduction will be much smaller. Teams should first test the cache hit rate before deciding whether to migrate, rather than relying solely on the highest ratio shown in marketing materials.
Data retention is another cost factor. By default, using Fable, companies must accept a 30-day retention period for security monitoring purposes. Enterprises that meet the conditions specified by Enterprise Frontier Safeguards will be able to store their activity data in their own cloud environments in the future and have it manually reviewed by their staff. Until this solution is fully implemented, eligible customers can temporarily opt for zero data retention. There are time and qualification restrictions in place; it cannot be assumed that all companies have already obtained the ability to manage their logs autonomously.
The baseline description for Anthropic also highlights some comparison pitfalls. Fable was tested with production protection enabled at version 5.1; after some security tasks triggered downgrades, they were either completed by zero or by other models. The task files for OSWorld were also updated in August 2026, so the results of the old versions cannot be directly compared. A more practical approach for buyers is to rerun their own code and tools with the longest tasks, and then to track the completion rate, the number of manual interventions, the cache hit rate, and the proportion of secure routes used.
The significance of 5.1 is not to make a single conversation more capable of answering questions, but to integrate the model into longer, more expensive, and more operationally risky workflows. Caching price reductions address the issue of repetitive contexts in billing, while hierarchical models solve part of the problem with limited capability exposure. Default retention and secure routing bring about new governance requirements. Truly mature deployments do not consider being "operable for several days" as a reason for unmonitored operation; instead, they include stage checkpoints, budget limits, permission boundaries, and model identities in the task records.
