The Practical Guide to Claude Fable 5
The most important thing about Fable 5 is not the benchmark. It is the production architecture you need around it.
What This Guide Covers
What Fable 5 actually is — the Mythos-class tier, and the “one model, two products” split with Mythos 5.
The specs that matter — context, output, modalities, and the always-on thinking model.
The money — real pricing, caching, batch, and the only cost metric that matters.
The safeguards classifier — refusals, the three fallback paths, and the billing trap. This is the section you cannot skip.
Adaptive thinking, effort, and task budgets — how to control depth and spend.
The agentic toolkit — memory, context editing, compaction, and the 1M window.
When to use it, and when you’re just burning money.
How to read its benchmarks like an engineer, not a fan.
How to run it in production — access, usage credits, rate limits, data retention.
The access situation, honestly — the suspension, the restoration, and why you should not build a single point of failure on it.
Most model-launch coverage is a benchmark screenshot and a vibe.
This is not that.
Claude Fable 5 is the most capable model Anthropic has released to the public, and if you are building on it, the interesting part is not the leaderboard.
It is that Fable 5 introduces a genuinely new API behavior (a model that can decline a request and force you to fall back) and a price tag that punishes anyone who reaches for it by reflex.
This is a manual on what it is, what it costs, the one behavior that will break your integration if you ignore it, when to actually use it, and how to run it without lighting money on fire.
What Is Fable 5
Anthropic introduced a new capability tier above its Opus class, called Mythos.
Fable 5 is the first Mythos-class model made generally available.
Its API ID is claude-fable-5, and Anthropic describes it as its most capable widely released model, built for the most demanding reasoning and long-horizon agentic work.
Here is the framing that makes everything else make sense: Fable 5 is one model sold as two products.
Its sibling, Claude Mythos 5 (claude-mythos-5), is the same underlying model with the safety classifiers removed.
Mythos 5 is not generally available. It is restricted to approved customers through Project Glasswing, Anthropic’s trusted-access program.
Fable 5 is the version everyone else gets: identical capabilities, wrapped in classifiers that can decline high-risk requests.
That split is not marketing.
Anthropic held the previous Mythos-class model, Mythos Preview, back from public release entirely, citing offensive cybersecurity capability that outran its safeguards — reportedly finding zero-day vulnerabilities autonomously at a scale prior models could not touch.
Fable 5 is how that capability reaches the public: not by loosening the model, but by bolting on a classifier layer and routing the dangerous categories elsewhere.
So when you call Fable 5, you are calling a frontier model with a bouncer attached. The bouncer is the part that changes how you build.
The Specs That Matter
The numbers, from Anthropic’s own documentation:
Context window: 1 million tokens by default. Notably, the full window is billed at the standard per-token rate — a 900k-token request costs the same per token as a 9k-token one, with no long-context surcharge tier.
Output: up to 128k tokens per request.
Modalities: text, image, and file input; text output. Vision is supported at launch.
Capabilities: tool calling, prompt caching, web search, structured outputs (JSON schema), computer use, code execution, and a memory tool.
Thinking: adaptive thinking, always on. There is no “thinking off” switch, and the raw chain of thought is never returned. More on both below, because both have practical consequences.
Data retention: 30-day retention. Fable 5 and Mythos 5 are designated Covered Models and are not available under zero-data-retention. If your compliance posture requires ZDR, Fable 5 is off the table today — know that before you design around it.
The Money: Pricing and Real Unit Economics
Fable 5 lists at $10 per million input tokens and $50 per million output tokens. That is exactly double Claude Opus 4.8’s $5/$25, and it is the highest rate Anthropic charges for a widely available model.
The 5:1 output-to-input ratio is standard — generating tokens costs more than reading them.
The levers that actually move your bill:
Prompt caching drops cache reads to roughly $1 per million tokens — a 90% discount on repeated input. Cache writes run about $12.50/MTok for the 5-minute window and $20/MTok for the 1-hour window. For any agentic workload with a stable system prompt, tool catalog, or codebase scaffolding, this is the single most effective spend lever you have. (We have written a whole piece on why; it applies here with interest.)
The Batch API halves both rates ($5/$25) for work that is not time-sensitive.
Make the cost concrete.
A single agentic turn that reads 50k tokens of context and writes 10k of output runs about $1.00.
A heavy planning pass (200k in, 40k out) is about $4.00. Run that in a multi-turn autonomous loop for a few hours and the meter is real.
The Cost-Per-Task Table
Fable 5 pricing only becomes real when you convert it from tokens into tasks.
A million-token price looks abstract on a pricing page, but it becomes very concrete when an autonomous workflow starts reading large context, writing long outputs, retrying, and looping across multiple turns.
The table below uses the base API pricing of $10 per million input tokens and $50 per million output tokens. It does not include prompt caching, batch discounts, cloud-provider premiums, or fallback behavior.
Treat it as a simple first-pass estimate for understanding the shape of the bill.
