Claude Fable 5 and Claude Mythos 5 challenge the habit of treating a model name as a complete product description. Anthropic says the two offerings use the same underlying model, with the same capability, context capacity, maximum output, and token price. Their user experiences differ because one adds deployment classifiers for broad use and the other is available only through a tightly controlled research program.

Anthropic launched both on June 9, 2026. Fable became its most capable broadly available model, while Mythos was offered to a limited set of approved organizations under Project Glasswing. Access was paused three days later under unusual government restrictions, then restored on July 1 after those restrictions were lifted and Anthropic updated Fable’s classifiers.

The episode is best understood as a combination of model deployment, safety policy, access control, and export regulation—not as a normal outage or a simple “unsafe model recalled” story.

What the two models share

Fable 5 and Mythos 5 provide a one-million-token context window and maximum 128,000-token output. Both cost $10 per million input tokens and $50 per million output tokens. Both use always-on adaptive thinking that cannot be disabled.

Anthropic also requires at least 30 days of data retention for the models and does not support Zero Data Retention. That is a significant constraint for organizations handling sensitive information. Teams should verify contractual, regulatory, and internal retention requirements before sending data to either offering.

“Same underlying model” means Mythos is not a larger hidden checkpoint and Fable is not a smaller capability tier. It does not mean every request produces the same answer. Deployment classifiers can refuse a request or send it through a different model, so the effective response path may differ even when the initial model is shared.

Fable’s classifier layer

Fable is designed for broad access and includes classifiers around high-risk domains. Anthropic’s launch material describes controls for areas such as cybersecurity, biology and chemistry, and attempts to extract or distill frontier model capabilities. Current developer documentation includes refusal categories such as cyber, bio, frontier_llm, and reasoning_extraction.

In the Messages API, a Fable classifier refusal is normally a successful HTTP 200 response with stop_reason: "refusal". Applications must inspect the response body. Treating every 200 as a completed answer can pass missing content into another system or trigger a confusing retry.

A refusal does not prove malicious intent. Defensive security, debugging, biological research, and model evaluation can resemble restricted behavior. The application should preserve context, explain the boundary where appropriate, and provide an authorized escalation path rather than asking the model repeatedly until wording slips through.

Anthropic reported at the initial launch that more than 95 percent of sessions did not require fallback. That was an early average observation, not a service-level agreement or a promise that any particular request had a 95 percent acceptance probability. The company later deployed stricter classifiers and acknowledged more false positives, so the launch statistic should not be projected onto the redeployed system.

Mythos is not an unrestricted public model

Mythos does not include the deployment classifiers used by Fable. It would still be misleading to call it “unsafeguarded,” “unlimited,” or a public jailbreak edition. Access remains invitation-only through Project Glasswing, and Anthropic approves a limited set of organizations for research that may require capabilities blocked by broad deployment controls.

Institutional access is itself a safeguard. Eligibility, contracts, monitoring, retention, and research purpose shape the environment even when the Fable classifier layer is absent. Mythos remains subject to Anthropic’s terms and other controls.

The distinction is useful for safety research. Broad consumer deployment needs low-friction protection at enormous scale. Qualified research sometimes needs to study behavior that those protections hide. Serving the same core model through two access paths allows Anthropic to separate those contexts without claiming one classifier policy is appropriate for every user.

How fallback actually works

Fable can be configured to hand a classifier-refused request to Opus 4.8, but this is not automatic in every API call. Developers must explicitly configure Anthropic’s server-side fallback beta, use SDK middleware, or implement their own handling.

Server-side fallback is triggered by safety-classifier refusal. It does not solve rate limits, overload, network failures, or server errors. Those require normal reliability logic.

The response should disclose the model that produced it, and applications should log the route. Otherwise a team may believe it evaluated Fable when many answers actually came from Opus. Fallback also changes cost, latency, and capability.

A sound contract distinguishes:

Consumer Claude products may implement routing differently from a direct API application. Do not infer the API contract from a product conversation or assume a product fallback exists in custom code.

The June suspension and July redeployment

The timeline prevents several common misreadings:

  1. June 9: Anthropic launched Fable 5 and Mythos 5.
  2. June 12: the US government imposed immediate export controls requiring restrictions on access by foreign nationals.
  3. June 12: Anthropic paused both offerings for all users because it said it could not reliably verify nationality in real time.
  4. June 30: Anthropic announced that the export controls had been lifted and access would resume the following day.
  5. July 1: Fable returned to first-party products globally, while Mythos returned only for approved organizations.

A report about bypassing Fable’s initial classifiers preceded the government action. Anthropic said the demonstrated behavior did not reveal unique Mythos-level capability and could be reproduced with weaker models. During the pause, the company trained updated classifiers and reported blocking more than 99 percent of the specific technique described in that report.

That number applies to a particular technique, not all possible bypasses. Anthropic also stated that the tighter classifier would increase false positives for legitimate coding and debugging. The redeployment therefore represents a tradeoff, not a claim of perfect immunity.

Describing the event as “Fable was recalled because it was hacked” loses the governing fact: the immediate suspension resulted from export controls and the inability to verify nationality. Describing it as “only politics” would also omit the classifier work Anthropic performed during the pause.

Choosing between the offerings

Most developers do not have a practical choice: Fable is the broad product, and Mythos is limited to approved Project Glasswing participants. Teams should design around Fable’s refusal contract unless they have documented Mythos access.

Fable fits complex, high-value agent and knowledge work where its price and latency are justified and the classifier profile is acceptable. Build refusal-aware user experiences, keep Opus fallback optional and visible, and evaluate legitimate false positives from real workloads.

Mythos may fit qualified research that needs to observe the underlying model without Fable’s deployment classifiers. It is not a shortcut around policies for ordinary production applications.

For both models, the high token price, mandatory adaptive thinking, and retention requirements demand cost and data governance. A million-token window can produce a very large bill and a large disclosure surface if used carelessly.

Safeguards around identity-changing media

Face swap and video character replacement illustrate why model-level safeguards are only one layer. Claude can interpret a request and coordinate an external MCP tool, but an allowed response does not establish that the user owns the footage, has consent from the depicted person, or may publish the result.

Medux offers separately configured workflows for image face swap with Claude MCP and video character replacement. Medux performs those transformations as an external service; it is not native to Fable or Mythos, and the language model choice does not grant media rights.

A responsible application should enforce safeguards outside Claude:

  1. collect explicit authorization for source and target identities;
  2. restrict supported use cases and reject deceptive or abusive purposes;
  3. show the exact files and requested transformation before upload;
  4. require human approval before the billable external task;
  5. store an audit record, consent reference, task ID, and output hash;
  6. label synthetic or materially altered media where required;
  7. require separate review before distribution.

If Fable refuses, the application should not automatically route to Mythos or repeatedly rephrase the request. A configured fallback should follow an approved policy and remain visible in the audit trail. For high-risk identity work, a classifier acceptance is only the beginning of authorization, not its conclusion.

Fable and Mythos show why AI governance cannot stop at base-model comparisons. The complete system includes who can access it, what classifiers run, how refusals and fallbacks behave, what data is retained, which external tools are available, and which human decisions remain mandatory.