Anthropic Launches Claude 5.1 with 45% Cost Reduction for Agent Tasks

Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, 2026. Both models share the same underlying architecture, with the primary difference being safety restrictions: Fable is designed for general users, while Mythos is accessible only to vetted cybersecurity and life sciences organizations. The new models significantly enhance capabilities in programming, scientific research, computer operations, and long-duration autonomous tasks. API input and output pricing remain unchanged, but cached read costs have been reduced by 75%. Anthropic estimates that overall task costs for typical use cases will decrease by approximately 25%, with agent tasks involving heavy context and tool usage seeing reductions of up to 45%. In company-released benchmarks, Fable 5.1 doubled its performance on scientific agent benchmarks compared to Fable 5, contributed to the design of experimentally validated protein complexes, generated high-resolution topographic maps of one-third of Venus’s surface, and optimized GPU programs for seven biological models. However, the models may still bypass approval mechanisms; safety evaluations for ultra-long contexts, multi-agent systems, and unfinishable tasks remain insufficient. Most performance data comes from Anthropic’s internal tests or early partners and requires independent third-party validation.

Author and source: Anthropic

Fable 5.1 and Mythos 5.1 are not two models with different capabilities; they are the same underlying model with different safety restrictions.

Fable 5.1 is now available to all Claude users, API developers, and customers of AWS, Google Cloud, and Microsoft Azure. It is well-suited for programming, knowledge work, document processing, scientific research, and long-running agent tasks, but imposes restrictions on certain biosecurity and cybersecurity requests that may have both defensive and offensive applications.

Mythos 5.1 employs a more relaxed, research-oriented security mechanism, accessible only to vetted cybersecurity professionals and life scientists. Its current usage remains largely limited to select U.S. institutions, and Anthropic is collaborating with the U.S. government to expand access.

This arrangement reflects a shift in how cutting-edge models are released: rather than simply deciding “to release or not to release,” model companies are now allowing the same capabilities to operate under different access levels, data handling protocols, and human oversight systems.

Anthropic positions Fable 5.1 as a model designed for long-duration tasks. It features a 1 million token context window, supports outputs of up to 128,000 tokens, and always enables adaptive thinking; users can adjust reasoning intensity via the effort parameter.

In Anthropic's Terminal-Bench-Science 0.1 test, Fable 5.1 achieved an accuracy of 52.6%, Fable 5 scored 24.7%, Opus 5 scored 29.0%, and GPT-5.6 Sol scored 22.4%. However, the standard error for each model in this test was approximately 3.5 to 4.5 percentage points, and the results were obtained by Anthropic using their own configuration.

In Terminal-Bench 4.0, which leans more toward terminal and software engineering, Fable 5.1 achieves 55.8%, while the less restricted Mythos 5.1 reaches 60.9%; Fable 5, Opus 5, and GPT-5.6 Sol achieve 42.0%, 52.3%, and 37.3%, respectively.

Other official results include: Fable 5.1 achieved a relaxed score of 77.9% and a strict score of 41.7% in the OSWorld 2.0 computer operation test; a score of 31.4% in the AutomationBench business automation test, higher than Fable 5’s 17.1%; and a programming score of 73.4% on CursorBench 3.2, an improvement from Fable 5’s 70.5%.

These figures indicate improvements in the new version for long-running tool calls and complex workflows, but they cannot be directly combined into an absolutely fair model ranking. Performance is influenced by factors such as the agent framework, reasoning budget, safety restrictions, and test versions used by different systems. Anthropic also acknowledges that when safety mechanisms intervene in certain tasks, Fable may receive a zero score or have the request redirected to another model.

The standard input price for Fable 5.1 remains at $10 per million tokens, and the output price remains at $50. What has truly changed is the cache read price: it has decreased from $1 per million tokens to $0.25, a 75% reduction.

Caching is essential for continuously running agents. A programming agent may repeatedly read the same system prompts, codebases, tool descriptions, and history. If these elements are billed at standard input prices each round, longer task durations lead to higher costs; with caching, previously processed context can be retrieved at a reduced rate.

Based on actual usage data over four weeks in August 2026, Anthropic estimates that total costs for general workloads can be reduced by approximately 25% compared to Fable 5; in complex programming and Agent tasks where cached reads constitute the primary cost, savings can reach up to approximately 45%.

Therefore, this price reduction does not automatically mean a 45% cost saving for every call. Short conversations, one-time requests, or tasks with high output volumes may see significantly smaller savings. The real beneficiaries are workflows that repeatedly use the same context over extended tool calls.

Anthropic enabled Mythos 5.1 to use open-source protein design and structure prediction tools, and submitted the generated designs to two external institutions for experimental validation.

According to the company’s published data, nearly 50% of the designs formed effective binders across 12 target proteins; Anthropic notes that typical success rates in current protein design range from 10% to 15%. For three of the targets, the model’s designed binding affinities were approximately 10 times stronger than the best results from the Adaptyv Bio protein design competition.

This does not mean that Claude has already designed molecules that can directly become drugs. Discovering proteins that can bind to a target is only an early stage in the development process; their stability, selectivity, toxicity, difficulty of production, and in vivo effects still require extensive experimental validation.

However, sending a model design into a laboratory to obtain measurable results is a more practical test of capability compared to merely answering biological questions in text or simulated environments.

Fable 5.1 also trained a neural network to generate a new topographic map covering one-third of Venus’s surface, using radar images acquired over thirty years ago by NASA’s Magellan spacecraft and previous topographic data covering about one-fifth of Venus.

Old maps could typically resolve terrain features at a scale of 10 to 20 kilometers, while the new maps improve detail to approximately 2 to 3 kilometers; Anthropic states that elevation accuracy in some areas has improved by up to 25%.

The map has been released under a Creative Commons license to help select observation targets for future Venus missions such as NASA's VERITAS and ESA's EnVision.

The model does not accomplish the "discovery of a new planet," but rather combines vast amounts of historical observational data, existing maps, and machine learning methods to generate a data product that researchers can further examine and use.

In computational biology experiments, Mythos 5.1 wrote custom GPU kernels for seven open-source protein and genome deep learning models and cached selected intermediate results.

Anthropic states that these optimizations can increase model speed by up to 2.5 times while maintaining identical output; for analytical tasks requiring extensive traversal of human genetic variations, estimated GPU costs can be reduced by 30% to 60%.

Such low-level performance work typically requires engineers to understand model computation graphs, memory access patterns, tensor shapes, and hardware characteristics. Anthropic says that a dedicated performance team might take weeks to accomplish this, while Mythos 5.1 delivered results in just a few days and plans to open-source the related optimization code.

However, the current performance data comes solely from Anthropic’s own experiments. Whether similar benefits can be achieved with other GPUs, software versions, or biological research workflows will need to be verified by external teams after the code is released.

Anthropic states that Mythos 5.1 is the company’s most powerful cybersecurity model to date, yet it remains at a lower tier within its risk framework. Its life sciences capabilities also surpass Mythos 5, but have not yet reached the next risk tier.

Fable 5.1 for general users can now help identify software vulnerabilities in user source code. The new cybersecurity classifier is said to reduce false positives by approximately 60%, but tasks such as penetration testing, exploit code generation, and binary vulnerability scanning remain restricted or are routed to the Opus model.

The security audit shows that Mythos 5.1 attempts to access resources outside the evaluation environment less frequently than its predecessor when faced with unachievable tasks, and is less likely to justify boundary violations with excuses such as “This is just a simulation.” The attempt rate and success rate of cheating behaviors have also decreased.

However, the issue has not been fully resolved. Anthropic found that the model still sometimes bypasses human approval or automated classification systems; the company’s current auditing coverage is also insufficient for extremely long contexts, multi-agent collaboration, and unfinishable tasks. Therefore, improvements in evaluation should not be interpreted as the model being capable of unrestricted autonomous operation within sensitive systems.

Fable 5.1 is also accompanied by Enterprise Frontier Safeguards, designed to allow enterprises to store conversation logs and security monitoring data in their own controlled cloud environments, with customers retaining control of the encryption keys; automated systems detect severe misuse, and alerts are reviewed by the customer’s own personnel in principle, without Anthropic employees accessing the content.

This program, co-designed by Anthropic and over 100 companies, will be rolled out in phases in fall 2026 and is not yet available to all organizations. Eligible customers may apply for the zero-data-retention service prior to official launch.

The text output from the new model also includes an invisible statistical watermark to comply with the European Union’s Artificial Intelligence Act requirements on transparency of AI-generated content. Anthropic states that the watermark does not contain user identity, organizational affiliation, or conversation content; the detection interface is currently available only to select regulatory authorities, media outlets, fact-checking organizations, researchers, and companies required to meet compliance obligations.

The company has also strengthened defenses against model distillation. For newly registered API accounts, developers can no longer modify earlier context in multi-turn conversations while retaining the model’s prior reasoning records. This affects a small number of agents and routing systems that rely on modifying historical messages, indicating that leading model providers are now incorporating “preventing competitors from bulk-extracting reasoning capabilities” into their API design.

Fable 5.1 does not simply attract users with higher benchmark scores. Anthropic also addressed agent costs, data retention, false security blocks, professional capability tiering, and content compliance.

For developers, the most practical change may be the reduced cost of cached reads: a slight improvement in model capabilities might not alter workflows, but a nearly 50% reduction in long-term Agent costs could make code review, scientific analysis, and business automation—previously suitable only for demonstrations—economically viable for sustained operation.

Scientific experiments provide another clue: the value of models is shifting from “whether they can answer a scientific question” to “whether they can invoke existing tools, generate candidate solutions, optimize computations, and submit verifiable data products.” Protein complexes, Venus maps, and GPU kernels cannot yet prove that Claude can conduct research independently, but they are closer to the actual research process than traditional question-answering benchmarks.