Claude Fable 5: the new flagship, the new price, and the first model that can refuse you

Status update — July 1, 2026: Fable 5 is back. The US export-control directive that forced this model offline was lifted on June 30, and Anthropic restored access to both Claude Fable 5 and Claude Mythos 5 on July 1, 2026. claude-fable-5 is callable again on the Claude API and across subscription plans. The suspension lasted three weeks, from June 12 to July 1. See Anthropic's statement. Two things are worth carrying forward from the outage: keep the fallback path you built (the fallbacks parameter now supports a "default" mode that picks fallback models for you), and read what to do when your AI model disappears — this will happen again to some model, somewhere. Note also that Fable 5 is no longer the top of the lineup on price-performance: Claude Opus 5 launched July 24, 2026 at $5/$25 — half the price of Fable 5's $10/$50 — and Anthropic positions it as close to Fable 5's frontier intelligence.

---

Claude Fable 5 (claude-fable-5) shipped on June 9, 2026 as Anthropic's most capable widely released model. It's the public, generally available version of the Mythos model class — the same model line that, until now, only Project Glasswing partners and critical-infrastructure operators could touch.

At launch it was generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry — until the June 12 directive pulled it from all of them.

Two things make this release different from a routine model bump: the price doubled, and Fable 5 is the first Claude model that can refuse a request in the middle of an API call. Both have direct consequences for anyone with Claude in production.

The specs

	Claude Fable 5	Claude Opus 4.8 (for comparison)
Model ID	claude-fable-5	claude-opus-4-8
Input price	$10 / million tokens	$5 / million tokens
Output price	$50 / million tokens	$25 / million tokens
Context window	1M tokens (default)	1M tokens (default)
Max output	128k tokens	128k tokens
Thinking	Adaptive, always on	Adaptive (effort defaults to high)
Safety classifiers	Yes — can refuse	No

Fable 5 is twice the price of Opus 4.8 on both input and output. That alone should make you deliberate about where you point it. More on that below.

One more cost wrinkle: Fable 5 uses the tokenizer introduced with Opus 4.7, so the same text produces roughly 30% more tokens than models before 4.7. If you're budgeting off old token counts, re-measure with the token counting API using model: "claude-fable-5" before you trust any cost estimate.

What "the model can refuse you" actually means

On Fable 5, Anthropic runs safety classifiers on the incoming request and again during generation. When a classifier declines a request — primarily in high-risk cyber, biology, and chemistry domains, plus attempts to reverse-engineer the model — the Messages API does not throw an error. It returns a normal HTTP 200 response with:

{
  "stop_reason": "refusal",
  "stop_details": {
    "category": "cyber",
    "explanation": "..."
  }
}

stop_details.category is "cyber", "bio", or — new on Fable 5 — "reasoning_extraction" (blocked under the Terms of Service rules on duplicating model outputs). No beta header is required to read it.

The trap: if your code branches on stop_reason === "end_turn" or just reads response.content[0].text, a refusal will surface as an empty or malformed result rather than an error your monitoring catches. You have to handle "refusal" explicitly.

Billing detail that works in your favor: you are not charged for a request that is refused before any output is generated.

Handle the refusal, then fall back

A request Fable 5 refuses can usually be served by another Claude model — Opus 4.8 has no classifiers. The simplest, most portable pattern is a client-side check and retry:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

async function ask(prompt: string) {
  const res = await client.messages.create({
    model: "claude-fable-5",
    max_tokens: 4096,
    messages: [{ role: "user", content: prompt }],
  });

  if (res.stop_reason === "refusal") {
    // res.stop_details.category -> "cyber" | "bio" | "reasoning_extraction"
    // Retry on a model without the classifiers.
    return client.messages.create({
      model: "claude-opus-4-8",
      max_tokens: 4096,
      messages: [{ role: "user", content: prompt }],
    });
  }

  return res;
}

Anthropic also ships two managed paths so you don't hand-roll the retry:

• Server-side fallback — an opt-in fallbacks parameter (beta on the Claude API and Claude Platform on AWS; not supported on the Message Batches API) re-runs a refused request on a model you name, billed at that model's rates.
• Client-side fallback — official SDK middleware (TypeScript, Python, Go, Java, C#) that retries from the client on any platform.

Either way, fallback credit refunds the prompt-cache cost of switching models, so you don't pay the cache penalty twice. The full mechanics are in Refusals and fallback.

Other breaking differences from Opus 4.8

These are specific to Fable 5 (and Mythos 5). The Messages API is unchanged for Opus, Sonnet, and Haiku.

• Adaptive thinking is the only thinking mode. thinking: {"type": "disabled"} is not supported. Manual extended-thinking budgets and assistant prefill both return a 400 error. Control reasoning depth with the effort parameter instead.
• Raw chain of thought is never returned. thinking.display defaults to "omitted"; set "summarized" if you want readable reasoning summaries. Pass thinking blocks back unchanged in multi-turn conversations on the same model.
• No zero data retention. Fable 5 requires 30-day data retention and is a designated Covered Model — it is not available under ZDR. If your contract mandates ZDR, you can't use Fable 5 yet; stay on Opus 4.8.

If you're coming from the old Mythos Preview, the migration is mostly a model-name swap to claude-mythos-5; see the migration guide.

What Claude Mythos 5 is

claude-mythos-5 is the same underlying model as Fable 5, with the safety classifiers removed. It never refuses. It is not generally available — it stays locked to approved Project Glasswing participants: vetted cyber defenders and critical-infrastructure operators doing exactly the high-risk work Fable 5 blocks. If you don't have Glasswing access, Fable 5 gives you the same capabilities everywhere except those restricted domains.

Is it actually better?

Anthropic reports Fable 5 scores more than 10% higher than Opus 4.8 on some benchmarks, with strength in software engineering, knowledge work, and vision. Early third-party testing backs the direction: analytics platform Hex reported Fable as the first model to clear 90% on its core benchmark, and app builders cited stronger one-shot app generation and tool-calling than any prior model.

Treat those as a real step up at the top end — not a reason to move everything. The honest version: Fable 5 is the model for the hardest reasoning and longest-horizon agentic work, where a 10% quality gain is worth paying double for and the occasional refusal-and-fallback is acceptable. For the large majority of production traffic — drafting, summarizing, classification, routine tool use — Opus 4.8 at half the price, with no refusal handling required, is still the right default. Route to Fable 5 deliberately, not globally.

Availability and the pricing clock

Note: the free-access window below (through June 22) has passed; Fable 5 now bills at usage credits on subscription plans and at $10/$50 on the API. Access was restored on July 1, 2026 after the June 12–30 government suspension (see the status note at the top).

As launched, Fable 5 was free in Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026, then moving to usage credits on those plans, and billed at $10/$50 on the API from day one.

If you're putting Fable 5 in an app, the API price is what matters — and that, plus the lesson the suspension just taught everyone about a single model vanishing overnight, is exactly why the "route deliberately, keep a fallback" advice earns its keep.

What to do this week

1. Find every place you'd swap the model string. For each, add an explicit stop_reason === "refusal" branch with a fallback to Opus 4.8.
2. Re-measure token counts under the new tokenizer before trusting any cost projection.
3. Decide the routing rule. Default to Opus 4.8; send only your hardest tasks to Fable 5. Write that rule down so it doesn't drift into "Fable everywhere" by habit.
4. Check your ZDR requirement. If you're contractually on zero data retention, Fable 5 is off the table for now.

Related reading

• Claude Opus 4.8: the new default — the model Fable 5 sits above, and your fallback target
• Choosing the right Claude model — the routing framework, now with a fourth tier
• Claude cost optimization — why "route deliberately" is a cost decision, not just a quality one

---

Source: Introducing Claude Fable 5 and Claude Mythos 5 and the Claude Platform release notes, June 9, 2026.