Claude Fable 5 · 2026 Edition

Claude Fable 5

Power User Guide

NEW

Anthropic's most capable public model — the first Mythos-class model made safe for everyone. Run it for days, let it plan across stages, and check its own work. Here's how to get the most out of it.

Guide by Anastasiia Shapovalova claude-fable-5 Mythos-class 5th generation Anastasiia Shapovalova

What is Fable 5?

The top tier of the Claude 5 lineup — raw intelligence plus reliable, autonomous action.

Mythos-class, made general. Same underlying model as Mythos 5, released with strong safeguards so anyone can use it.
Long-horizon by design. Sustains complex, asynchronous tasks earlier models couldn't finish.
Thorough & proactive. Plans its approach, tests as it goes, and refines toward the goal.
State of the art on coding, knowledge work, vision & computer use — its lead grows with task length.

The one-line rule: if a task fails or comes back weak on a smaller model, hand it to Fable 5 before you conclude it can't be done.

The specs

What you're working with under the hood.

Context

tokens in, by default

128k

Output

tokens per request

$10/M

Input

90% cache discount

$50/M

Output

Mythos-class pricing

Text, images & documents in. Raw chain-of-thought is never returned — thinking comes back summarized or omitted.

What's new vs. Opus 4.8

Four capabilities that change how you actually work with the model.

Days-long autonomy

Runs in an agent harness (Claude Code, managed agents) for extended sessions with minimal oversight.
Breaks big goals into stages and delegates work to sub-agents.

Self-verification

Writes its own tests and checks outputs against the original goal.
Builds its own harnesses and evals, and updates its own skills from what it learns.

Advanced vision

Reads diagrams, charts & tables nested inside files and PDFs.
Uses vision to critique its own coding output against a design or spec.

Hand-off, not hand-holding

Review completed work instead of supervising every step.
High-fidelity implementation for large migrations & multi-day builds.

Work it can now sustain

The class of problems that were out of reach before.

Coding

Large codebase migrations & complex implementations.
Multi-day autonomous sessions that plan, build & verify.
Designs implemented at high fidelity, checked by vision.

Knowledge work

Deep research & analysis, handed off end-to-end.
Doc-heavy work in finance, legal, analytics.
Deliverables that arrive ready to review.

How it checks itself

Proactive self-verification is the headline behavior — lean into it.

Give it the goal, not the steps. It scopes and sequences on its own.
Ask it to write tests first, then implement against them.
Let it review its own work — request a self-critique pass before it calls a task done.
Point it at a source of truth — a design, spec or dataset it can verify against with vision.

Safeguards & the Opus 4.8 fallback

Fable 5 ships with the strongest safeguards Anthropic has applied to any model.

Your request

Coding, research, knowledge work, vision…

→

Fable 5

Answers directly for the vast majority of sessions.

→

Routes to Opus 4.8

If a cyber / bio / chem / health classifier flags it.

You aren't charged Fable prices for a rerouted request.
Tuned conservatively — occasionally catches harmless prompts, but triggers in under ~5% of sessions on average.
Landed on Opus 4.8 for a harmless request? Rephrase or split the task and continue.

Where you can run it

Available to Pro, Max, Team & Enterprise.

✓Claude.ai — web, desktop & mobile chat.

✓Claude Code — agentic coding in terminal or IDE.

✓Claude Cowork — agentic knowledge work.

✓Platform / API — native, plus Bedrock, Google Cloud & Foundry.

model: "claude-fable-5"
context: 1M in · 128k out

When to use what

Fable 5 isn't always the answer — match the model to the job.

Reach for	When the job is…	Why
Fable 5	Ambitious, long-running & asynchronous — big projects to break down, research, build & verify.	Sustained reasoning + tool orchestration across many steps; its lead grows with task length.
Sonnet 5	Straightforward tasks, or apps making many fast model calls.	Faster & cheaper; frontier performance at scale for everyday work.
Opus 4.8	Hard single-shot problems, or anything a safeguard routes away from Fable.	Powerful general model; the built-in fallback for flagged requests.
Haiku 4.5	High-volume, latency-sensitive, lightweight tasks.	Fastest & lowest cost when deep reasoning isn't needed.

Power-user tips

Get more out of every long-horizon run.

Front-load context. The 1M window rewards full specs & examples up front.
Define "done." State acceptance criteria so it can self-check.
Let it plan first. Approve a staged plan, then let it execute.

Feed it visuals. Drop in PDFs & dashboards — it reads nested charts.
Cache your context. The 90% cache discount pays off on long prompts.
Batch the hand-off. Give a whole project, then review the result.

The Claude 5 family

Four tiers — pick the altitude your task needs.

Mythos — top tier: Fable 5 (safeguarded, public) & Mythos 5 (limited).
Opus 4.8 — powerful general model & Fable's safety fallback.
Sonnet 5 — frontier performance at scale, faster & cheaper.
Haiku 4.5 — fastest & leanest, for high-volume work.

The success formula

Clear goal

Full context

Right harness

Let it verify

Work you review, not run