Claude Fable 5: Features, Pricing, and Fallbacks

Fable 5 and Mythos 5 are the same underlying model. The difference is the safeguards. Mythos 5 runs without the safety classifiers and is offered only in limited release through Project Glasswing, mostly to cyber defenders and critical infrastructure providers. Fable 5 is the version Anthropic made safe for everyone, and it is the one you reach for on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

The naming follows a theme. A Mythos class model sits a tier above the Opus class in raw capability. Fable comes from the Latin fabula, which means a thing that is told, close in spirit to the Greek mythos. The safeguards are the only thing that separates the two names, and that distinction is the whole point of releasing Fable to the public.

For most of this post the practical takeaway is simple. When you call claude-fable-5 and your request is not flagged, you get the full Mythos class model. Anthropic's own data shows more than 95% of Fable sessions involve no fallback at all, so for the large majority of work you are talking to the strongest model the company has put in general hands.

Fable 5 costs twice as much per token as Opus 4.8. That premium buys you the strongest reasoning and the longest autonomous runs Anthropic has shipped broadly, but it also means cost control matters more, not less. If your workload is mostly routine, routing only the hard tasks to Fable and keeping the rest on Opus or Sonnet is the obvious move. We dug into that kind of tradeoff in our guide to cutting LLM inference cost, and the same logic applies here.

The 1M token context window is on by default, so you do not opt into it. The 128k output ceiling is generous enough to return whole files, long migrations, or multi document analyses in a single response, which is exactly the kind of long horizon work the model is built for.

A Mythos class model has frontier strength in areas like cybersecurity and biology, which is exactly what makes it dangerous in the wrong hands. So Anthropic put a separate set of classifier models in front of Fable 5. When a request looks like it touches a high risk area, the main model never answers. The response is handled by Claude Opus 4.8 instead, and the user is told the fallback happened.

The safeguards are deliberately conservative, so they sometimes catch harmless requests. Anthropic says they trigger in under 5% of sessions on average and plans to reduce false positives after launch. For builders, the takeaway is to expect the occasional fallback on benign cyber or bio adjacent prompts and to handle it gracefully rather than treat it as an error.

The biggest one is refusals. When Fable 5 declines a request, the Messages API does not throw. It returns a successful HTTP 200 with stop_reason set to refusal, and it tells you which classifier declined. You are not billed for a request that is refused before any output is generated, so handling the refusal is about user experience, not cost recovery.

// A refused request comes back as a normal 200 response
{
  "type": "message",
  "role": "assistant",
  "stop_reason": "refusal",
  "content": [],
  "usage": { "input_tokens": 412, "output_tokens": 0 }
}

// Opt into a server side retry on another model with the fallbacks param
{
  "model": "claude-fable-5",
  "fallbacks": ["claude-opus-4-8"],
  "messages": [ /* ... */ ]
}

A request Fable refuses can usually be served by another model. You can pass the fallbacks parameter to have the API retry for you, or use the SDK middleware to retry from the client on any platform. Fallback credit refunds the prompt cache cost of switching, so a retry does not make you pay the cache cost twice.

If you are weighing Fable against the rest of the field before you wire it in, our comparison of the major LLM providers frames where each one fits in an agentic stack.

Pricing is $10 per million input tokens and $50 per million output tokens, the same for both Fable 5 and Mythos 5. On the Claude API and consumption based Enterprise plans, Fable 5 is fully available now. For subscription plans the rollout is more cautious. From launch through a short included window, Fable 5 comes with Pro, Max, Team, and seat based Enterprise plans at no extra cost. After that, using it draws on usage credits until capacity allows Anthropic to fold it back into the standard plans.

Mythos 5 is not generally available. It is restricted to approved Glasswing partners with cyber safeguards lifted, and soon to a small set of biology researchers through a trusted access program. If you do not have Mythos access, Fable 5 is the Mythos class model you use, and for nearly every request it behaves identically.

If you are deciding how a model this capable fits into a real product rather than a benchmark, that is a systems question more than a model question. It is the kind of work we cover in our agentic AI consulting engagements, where the model is one component inside the orchestration, memory, and safety layers around it.

The primary sources for everything above are Anthropic's launch announcement and the Claude API documentation.