2026.07.29 2026.07.29

What Is Claude Fable 5? Inside the Mythos Class That Now Sits Above Opus

swiftwand

[Editor’s note, July 25, 2026] After this article was finalized, Anthropic released Claude Opus 5 on July 24 — same $5/$25 API pricing as Opus 4.8, 1M context, 128K output, and the new default on the Max plan. The facts below are accurate as of mid-July 2026, but revisit any Fable-5-versus-Opus decision with Opus 5 in the picture. A full review is coming soon.

When the name Claude Fable 5 first appeared, most people had the same questions. What is “Fable,” if not Opus or Sonnet? How does it relate to the simultaneously announced Mythos 5? And should I be using it? The June 9, 2026 announcement was not just another model launch — it stacked a fourth floor, the Mythos class, on top of a hierarchy that had been stable at three: Haiku, Sonnet, Opus. This article dissects Claude Fable 5 using only primary sources — the official announcement and documentation — to settle what is new, how it works, and when to use it.

Why this is not “just another model”
What the Mythos class is — a four-tier hierarchy
The specs — 1M context, 128K output, /
Fable versus Mythos — same capability, different taps
Inside the safety classifiers — three categories and the shape of refusal
Recovering from refusals — three fallback routes and the billing rules
Reading the performance claims at the right distance
Adaptive thinking and the invisible chain of thought
Fable 5 or Opus 4.8 — when to pay double
Closing — an ace, not a daily driver

忍者AdMax

Why this is not “just another model”

Three statements in the announcement set it apart from routine launches. First, Anthropic explicitly calls this its most capable generally available model ever. Second, the same underlying model ships under two names — Fable 5 and Mythos 5 — to different audiences under different terms. Third, safety classifiers are built into the response path itself, with refusal and model-switching defined as part of the API contract. The headline is not raw capability but delivery design: who gets how much power, through what plumbing.

Even the naming carries intent. Haiku, Sonnet, and Opus form a size series drawn from poetry; raising the number signals a generation change. This time the name comes from outside that series entirely — a declaration that this is not the next step on the same staircase but a different landing altogether.

What the Mythos class is — a four-tier hierarchy

The Mythos class sits above the familiar three tiers, and unlike them it carries both a capability ceiling and delivery conditions. Two models currently belong to it: the generally available Claude Fable 5 and the restricted Claude Mythos 5. Their API model IDs are claude-fable-5 and claude-mythos-5. Official documentation states they share the same capability, specs, and price; the only difference is the presence of safety classifiers.

Why a new tier instead of “Opus 5”? A higher number only communicates more performance. A separate tier writes a boundary into the naming system: from this capability level upward, the terms of delivery change. Industry has long done the same — machines above a certain output, chemicals above a certain purity, get their own grade and their own rulebook.

The specs — 1M context, 128K output, /

The shared specs are clear-cut: a 1M-token context window by default, up to 128,000 output tokens per request, and API pricing of $10 per million input tokens and $50 per million output — exactly double Claude Opus 4.8’s $5/$25. The pricing itself is a message: this is not the model for everything, but the ace you play on selected hard problems.

The 1M default context is what you feel in long agentic sessions, where history, tool results, and references pile up. Anthropic says Fable 5 stays focused across millions of tokens of context and works autonomously longer than any previous Claude model. The 128K output ceiling matters too: long design documents or large diffs that used to require chunked instructions can now come out in one pass. But capacity has a bill — filling the 1M window costs about $10 of input per request. Treat the window as headroom to use when needed, combined with prompt caching and context hygiene, not as something to fill by default.

Fable versus Mythos — same capability, different taps

Fable 5 is for everyone: generally available from day one on the Claude API, Amazon Bedrock, the Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Mythos 5 is not. It is delivered through Project Glasswing, a framework coordinated with the US government, to approved organizations in cyber defense and critical infrastructure, plus vetted biology researchers — access runs through Anthropic or cloud account teams.

Think of two taps drawing from the same source. The household tap has a filter; the licensed-operator tap opens fully. The water — the model’s capability — is identical. One condition applies to both: traffic is retained for 30 days, and zero-data-retention is not available. If you handle sensitive material, check that retention requirement first.

Inside the safety classifiers — three categories and the shape of refusal

The classifiers watch three categories: requests touching cyberattack capability, high-risk biology and chemistry, and attempts to extract model capability — distillation. All are dual-use domains with severe misuse potential. What is genuinely new is that refusal is specified API behavior: when a classifier declines a request, the Messages API returns a normal HTTP 200 with stop_reason set to refusal, including which classifier fired. Refusal is now part of the normal path you design for, not an exception.

In consumer products like claude.ai, flagged requests are answered by a fallback to Opus 4.8. Ordinary development, writing, and analysis will essentially never touch these classifiers — but if you work near security research or the life sciences, test how your questions classify before a deadline does it for you. The experience of a refusal is a quality step-down, not silence: the answer comes from one floor below.

Recovering from refusals — three fallback routes and the billing rules

Refused requests can usually be served by another Claude model, and the documentation defines three retry routes. Server-side fallback: attach a fallbacks parameter and the API retries on an alternative model automatically (beta on the Claude API and the Claude Platform on AWS). Client-side: build the retry with SDK middleware. Or fully manual on any platform. Billing is tidy: requests refused before output generation are not billed, and a fallback credit offsets the doubled prompt-cache cost of switching models. Refusal is designed not to be a loss.

Which route fits depends on your integration. The fallbacks parameter is the shortest path to safety; SDK middleware suits teams that want to own model choice and logging on retry; organizations with an existing multi-vendor abstraction fold it in manually. The important part is choosing one before Fable 5 reaches production — the same wiring doubles as your outage playbook, a point June made vividly.

Reading the performance claims at the right distance

The announcement says Fable 5 reaches the state of the art on nearly every capability benchmark tested, cites the top score among frontier models on Cognition’s FrontierCode — achieved even at medium effort — and highlights a customer story of Stripe compressing months of engineering into days on a 50-million-line code migration. These are vendor statements; our editorial policy keeps them at arm’s length until independent replications accumulate. Still, “top score at medium effort” matters if it holds: strength at shallow settings changes everyday cost-performance, not just contest results.

The structural facts carry more certainty than the numbers: a model with 1M context, 128K output, and long autonomous operation as an explicit design goal is now generally available. To test it yourself, dig three problems from your archive that a previous model failed, resubmit them unchanged, and count the round-trips. Your own failure archive is the only benchmark that reliably speaks for your workload.

Adaptive thinking and the invisible chain of thought

Two API-level differences from earlier models stand out. First, adaptive thinking is the only thinking mode: disabling the thinking parameter is not supported, and the model itself decides how much to reason based on the problem, while you set the ceiling through the effort parameter. The interface is no longer “think or not” but “how far may you think.”

Second, raw chains of thought are not returned. With thinking.display set to summarized you get a readable summary; with the default omitted you get empty thinking blocks. A natural split: summaries on during development to understand behavior, default in stable production. The documentation is explicit that both behaviors are specific to the Mythos class — Opus, Sonnet, and Haiku keep their existing Messages API behavior — so migration diffs concentrate on exactly these two points.

Losing the option to prevent thinking also shifts prompt design. Where you once toggled thinking to trade speed against quality, the Mythos class folds that trade into effort alone. Write prompts that state goals and constraints clearly and let the model allocate — that style matches this design.

Fable 5 or Opus 4.8 — when to pay double

The most common practical question: is Opus 4.8 enough? The skeleton answer: use Opus 4.8 where it suffices, and hand Fable 5 only the work where it does not. Three tests identify that work. Context volume — does the task hold hundreds of thousands of tokens of material at once? Autonomy length — must it run for hours holding a plan without intervention? Cost of failure — is one attempt expensive enough that success rate outweighs price?

Concretely: a typical read-100K-tokens, write-10K-tokens task costs roughly $1.50 on Fable 5 versus about $0.75 on Opus 4.8 at list prices. Run dozens of times daily, the gap compounds; for a few weighty decisions a week, the difference is easily repaid by success rate. Multiply frequency by stakes and decide. Conversely, there is no reason to default to Fable 5 for short interactive tasks, summaries, or routine code fixes — Opus 4.8 with effort tuning reaches all of it.

Closing — an ace, not a daily driver

Claude Fable 5 is the safety-valved tap that opens the Mythos class to the public: identical capability to Mythos 5, 1M context and 128K output, double Opus 4.8’s price, refusal as specified API behavior with fallback and billing relief built in. The operating conclusion: let Sonnet 5 and Opus 4.8 carry daily work, and reserve Fable 5 for jobs where giant context, long autonomy, and a high cost of failure coincide. An ace earns its value when you decide in advance where you will play it. For makers, most code-CAD iteration sits comfortably below this tier — but a long design session that must hold an entire mechanism’s constraints together is exactly where a context that never drops the thread starts to pay.

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】