Claude Fable 5 vs Claude Mythos 5

Anthropic announced Claude Fable 5 and Claude Mythos 5 on June 9, 2026. The company positioned Fable 5 as a Mythos-class model for general use and Mythos 5 as a limited-release model for selected cyberdefenders and infrastructure providers. [1] The official documentation adds a practical integration distinction: Fable can return a refusal path and fall back to another Claude model, while Mythos does not include the same safety classifiers. [2]

This comparison focuses on the facts that affect a buyer, developer, or platform team. It separates capability claims from availability, safeguards, billing, retention, API behavior, and operational fit. Anthropic’s own claims about benchmark performance and customer testing are identified as vendor-reported evidence rather than treated as independent proof.

For broader model and agent context, compare our coding-agent guide, multimodal AI analysis, and multi-agent protocol explainer.

What You'll Learn

• How Fable 5 and Mythos 5 differ in access, safeguards, and intended users
• What the official API documentation says about refusals, fallback, billing, and thinking behavior
• How to compare the models for software, research, enterprise, and sensitive workloads
• Which operational checks should be completed before moving a production integration

Claude Fable 5 vs Claude Mythos 5 at a Glance

The short version is that Fable 5 is the public model and Mythos 5 is the restricted model. Anthropic says they are the same underlying model and share the same specifications and pricing. The key separation is the safety layer. Fable 5 includes classifiers that can detect some high-risk requests and send them to a different handling path. Mythos 5 is available only to limited trusted users and has safeguards lifted in selected areas. [1] [2]

This table is a product distinction, not an independent ranking. The models may behave differently in a real workflow because safeguards change which requests are answered, redirected, or refused.

What Anthropic Means by Mythos-Class

Anthropic describes Mythos-class as a tier above its Opus class in capability. The company says Fable 5 is its most capable widely released model and is built for demanding reasoning and long-horizon agentic work. [1] These are Anthropic’s descriptions and should be read as vendor positioning unless supported by a specific evaluation or independent test.

The practical meaning of long-horizon work is that the model may handle more steps, more context, and more tool interaction before a human must intervene. That does not mean every long task will succeed. A long-running agent still needs a clear objective, access controls, checkpoints, logging, and a way to stop or reverse harmful actions.

Mythos 5 is not a separate public tier that every user can select. Anthropic says it is the same underlying model as Fable 5 with some safeguards lifted for a limited program. The product decision is therefore about trust, risk, and access governance as much as raw capability.

Where Fable 5 and Mythos 5 Are Available

Anthropic’s launch page says Fable 5 is available everywhere, while Mythos 5 is restricted to Glasswing partners and selected trusted programs. [1] The official documentation describes Mythos 5 as available through Project Glasswing and identifies it as the successor to Claude Mythos Preview. [2]

Availability can change after launch. The announcement records an interruption on June 12, 2026 and a redeployment update on July 1, 2026. [1] A team should therefore verify the current access status, region, account type, and plan terms instead of relying only on a launch article.

Access is also a governance decision. A general release can support normal product development with safety handling. A trusted program may support specialized cybersecurity or biology work but require screening, monitoring, retention terms, and a narrower set of approved users.

Safety Classifiers and Fallback Behavior

Fable 5 includes separate AI safety classifiers that detect potential misuse. Anthropic says that when the classifiers identify certain cybersecurity, biology, chemistry, or distillation requests, the request can be handled by Claude Opus 4.8 instead. The company says more than 95% of Fable sessions involve no fallback. [1]

The API documentation gives an important implementation detail. When Fable 5 declines a request, the Messages API returns stop_reason: "refusal" in a successful HTTP 200 response rather than returning an HTTP error. [2] That means a client should inspect the response body and stop reason instead of treating every 200 response as usable model output.

The documentation also describes a fallback parameter that can retry on another Claude model. A team should define whether fallback is acceptable for its task, whether the second model has the required context, and whether the result needs a new review step.

Capability Claims and Their Limits

Anthropic reports that Fable 5 and Mythos 5 can work autonomously for longer than previous Claude models. It cites early software-engineering testing, knowledge-work evaluations, vision tasks, memory experiments, and life-science research. [1]

The launch article includes specific vendor-reported examples. Stripe reported that Fable 5 completed a migration in a 50-million-line Ruby codebase in one day that the company said would otherwise have taken a team more than two months by hand. Anthropic also says internal Mythos 5 work accelerated aspects of a drug-design process by around 10 times and produced strong candidates for 9 of 14 protein targets under investigation. [1]

These examples are informative but not universal performance guarantees. They come from early testing, selected tasks, or Anthropic’s own research setting. A buyer should ask for the evaluation protocol, baseline, human involvement, tool access, success definition, and failure rate before translating a headline into a production plan.

Software Engineering Use Cases

Fable 5 is positioned for software engineering and long-running agentic work. The likely benefits are faster repository exploration, migration planning, test generation, code transformation, documentation, and multi-step issue work. Mythos 5 may be relevant to trusted teams working on high-risk infrastructure or cyberdefense where access to the less restricted model is justified.

Neither model removes the need for software engineering controls. A production workflow still needs code review, dependency checks, secrets protection, test coverage, deployment gates, audit logs, and rollback. A model that can make more changes per run can also increase the size of a failure if the tool permissions are too broad.

For model-generated images and content, our AI copyright guide covers a different but related verification problem. The common rule is to confirm what the system produced, how it was produced, and who is accountable for publication.

Pricing and Token Economics

Anthropic lists the same price for Fable 5 and Mythos 5: $10 per million input tokens and $50 per million output tokens. [1] The platform documentation repeats that the models share specifications and pricing. [2]

Price per token is not the same as cost per completed task. Long-horizon work can use more input context, more output, retries, tool calls, and human review. A lower token bill may be offset by failed runs or by the cost of validating a large change.

Teams should model cost per accepted result. Track input and output tokens, tool calls, retry rates, fallback frequency, review time, test failures, and the percentage of results that reach production. Compare the same task with the same acceptance criteria rather than comparing a short answer with a long agent run.

Thinking, Context, and Long-Running Work

The official documentation says Fable 5 and Mythos 5 always have thinking enabled. Passing thinking: {"type": "disabled"} is not supported. Developers can control thinking depth through the effort parameter. [2]

The documentation also says raw chain of thought is not returned. The thinking.display setting can return a readable summary or omit the thinking block. In multi-turn conversations, thinking blocks should be passed back unchanged when required by the model’s handling rules. [2]

These details matter for integrations that assume every model supports the same request shape. Update request validation, logging, retry logic, and prompt templates before switching model IDs. Long context and persistent memory can improve a task, but they also increase the amount of state that must be protected and reviewed.

Data Retention and Privacy

Anthropic says Fable 5 and Mythos 5 require 30-day retention for traffic on Mythos-class models and are not available under zero data retention. The company says the data is not used to train new Claude models or for non-safety purposes, and describes additional protections such as logging human access and deleting data after 30 days in almost all cases. [1]

The platform documentation labels both models as Covered Models and repeats that they are not available under zero data retention. [2] This should be part of a data-protection review, not an afterthought.

Before sending customer, employee, health, financial, or source-code data, confirm the applicable contract, retention period, access logging, deletion behavior, region, and incident process. If the terms do not fit the workload, use redacted data, a different service, or a non-production environment.

Fable 5 vs Mythos 5 for Different Users

Most teams should begin with the model they can access under their existing controls. Fable 5 is the practical choice for general product development because Anthropic describes it as broadly available and provides documented refusal and fallback behavior. Mythos 5 is for approved users whose work justifies a more restricted program and additional governance.

This is an operating recommendation based on the cited access and policy facts. It is not a claim that one model is universally better. Fit depends on the task, data, approval path, and ability to review output.

Migration Checklist for Claude API Teams

Before changing a production model ID, test the request and response contract. Confirm that the client accepts the model name claude-fable-5 or, where approved, claude-mythos-5. Check refusal handling, fallback behavior, thinking settings, context limits, token accounting, and error classification.

Run representative tasks rather than only benchmark prompts. Include long code changes, tool failures, ambiguous requirements, security-sensitive requests, and prompts that should be refused. Measure accepted output, review time, retries, cost per task, and user-visible errors.

Keep an escape route. Store the previous model configuration, preserve prompts and evaluation fixtures, and make rollback a documented operation. A model migration is complete only when the team can detect a bad result and return to a known configuration.

For broader phone, device, and software troubleshooting context, see our device performance guide and AI engineering role guide.

Bottom Line: Which Model Should You Use?

Choose Fable 5 when you need broad availability, the documented safety-classifier path, and a model suitable for general reasoning and long-running agentic work. Choose Mythos 5 only when your organization is eligible for the relevant trusted access program and has the governance needed for a less restricted model.

The most important difference is not the headline capability. It is the surrounding system. Fable 5 can decline or redirect some requests. Mythos 5 is more restricted in access and has selected safeguards lifted. Both still require evaluation, access control, logging, privacy review, and human accountability.

Anthropic’s official documentation gives teams enough information to begin a controlled comparison. It does not remove the need to test the exact prompts, tools, data, and failure modes used by your organization.