Guide

Claude Fable 5.1: how to prompt the model that shares Mythos's skills

Published 30 September 2026

A person at a wooden desk in a quiet library reading a long printed document with a pencil, a laptop open beside them

Short answerClaude Fable 5.1 is Anthropic's most capable model open to everyone, with the same capabilities as Mythos 5.1 plus safety filters. It thinks on every request, so raise or lower effort instead of saying "think harder", tell it what to leave out, and ask for the format you want because it uses less bold and fewer lists than older models.

Claude Fable 5.1 is the model to reach for when Claude Sonnet or Opus isn't enough. Anthropic released it on 1 September 2026 alongside Claude Mythos 5.1, which has the same capabilities but is limited to approved organisations. Fable is the version everyone can use.

Old prompts still work on it, but a few of its habits are different, and a small change in how you ask can make a big difference. This guide draws on Anthropic's Fable 5.1 prompting guide and release notes.

What Fable 5.1 is good at

Anthropic says the biggest gains over Fable 5 are in:

Long coding sessions. Multi-file features, large refactors and debugging that run for hours.
Documents, spreadsheets and slides. Taking an analysis from the first question to a finished document, a spreadsheet with live formulas, or a slide deck.
Research. Multi-step web research that follows up on what it finds.
Reading charts, tables and PDFs.
Long documents. It has a 1-million-token context window and can reply with up to 128,000 tokens.

It costs $10 per million input tokens and $50 per million output tokens, more than twice the price of Opus 5.5. For most jobs, Anthropic's guide suggests trying Opus first. See which Claude model to use.

The one thing to know about thinking

Fable 5.1 always thinks before it answers, and you can't turn that off. What you can change is the effort level. The default is high.

Asking it to "think step by step" or "think less" doesn't change how much it thinks. The effort setting does.
In the API, and in tools that expose it such as Claude Code, effort is a setting you choose. Check what your app offers.
Save the top levels for tasks where you've seen a real gain. Anthropic also notes that a new tool lets developers change effort in the middle of a conversation: high for the hard step, low for the routine ones.

6 habits to fix with your prompt

Anthropic lists several ways Fable 5.1 behaves differently from Fable 5. Here is what each means for you and what to say.

1. Its writing can be dense

Its sentences can be longer and have fewer paragraph breaks than you might like. Ask for what you want:

prompt

Write in short paragraphs of two or three sentences. Use plain words. No long sentences with more than one idea.

Anthropic's own short version for removing a mannered style is "Please remove all mannered prose."

2. It uses less bold, fewer headings and fewer lists

If you want structure, ask for it by name: "Put this in a table", "Use headings for each section", "Give me a bulleted checklist." If your saved prompts say things like "don't use bullet points", check whether you still need them. Anthropic suggests replacing blanket bans with a rule for when formatting is right: lists when the content is multi-part, plain prose in casual or emotional conversation.

3. It may copy from your source without quotation marks

When summarising a document, Fable 5.1 is more likely to reuse the source's exact words without marking them. If that matters (an essay, an article, a client report), say so:

prompt

Summarise the text below in your own words. If you copy any phrase exactly, put it in quotation marks and say where it comes from.

Then check any passage that sounds too polished.

4. It may answer from memory when you needed a search

At low effort, Anthropic says Fable 5.1 is more likely to answer from what it already knows instead of looking things up. For anything that changes (prices, rules, dates, availability) ask directly:

prompt

Search for the current answer before you reply. For prices, dates and rules, use up-to-date sources and give me the link for each.

5. It sometimes does more than you asked

Given an open-ended job, it may fix nearby things, add extras or write more tests than the change needs. Anthropic's fix is to say so plainly. A version you can use for any task:

prompt

Do only what I asked. If you notice something else worth doing, list it at the end as a suggestion. Don't do it unless I say so.

6. Long jobs can go quiet

On long tasks it may work for a while without saying anything. Ask for updates at points that suit you:

prompt

Before you start, tell me your plan in one line. While you work, give me a short update at each major step. At the end, write a recap that makes sense on its own: what you found, what you did and what's next.

Anthropic's own suggested wording for this is close to the last line: a closing recap "that stands on its own" so a reader who only sees the final message still has the full picture.

A copy-ready template for Fable 5.1

Fill in the brackets and paste it into Claude:

prompt — 6 blanks

You are [ROLE, e.g. a careful research analyst].

Task: [WHAT YOU WANT, in one or two sentences].
This is for: [WHO WILL USE IT AND WHY].
Done means: [WHAT A FINISHED RESULT INCLUDES].

Format: [SHAPE AND LENGTH, e.g. a table with 5 rows, then a 3-sentence conclusion].
Writing style: Short paragraphs, plain words.
Scope: Do only what I asked. Put extra ideas at the end as suggestions.
Accuracy: Use only the material below, plus a search for anything that may have changed. Mark exact copies with quotation marks. If something isn't stated, write "not stated".
Updates: Tell me your plan in one line first, then give a recap at the end.

<material>
[PASTE YOUR DOCUMENT, DATA OR NOTES HERE]
</material>

It works well with the research paper summary prompt, the executive summary prompt and the delegate a task prompt.

Things you might see

A refusal. Fable 5.1 has safety filters. If a request is blocked, the API returns stop_reason: "refusal" with a category. Anthropic says the filters give fewer false alarms than Fable 5 did at launch, and that finding vulnerabilities in source code is allowed. If a normal request is blocked, rewrite it to make the harmless purpose clear, or try a model such as Opus 5.5.
A watermark. Anthropic says text from Fable 5.1 and Mythos 5.1 carries a statistical watermark. It doesn't add visible characters or change the meaning or readability of the text. If you're a student, check your school's rules on AI use, because watermarks are one way AI text can be identified.

For developers: what changes in the API

Forced tool use isn't supported. tool_choice set to any or a named tool returns an error. Use auto, tell the model when to use a tool, and use strict tool use or structured outputs when you need valid JSON.
Keep the conversation append-only. Thinking blocks are tied to the conversation and the model. Editing earlier turns invalidates them. Add instructions with a mid-conversation system message rather than editing the system prompt.
Set effort deliberately. The default is high. Anthropic's notes suggest changing it per turn instead of holding one level for a whole session.
Cache reads are cheaper. $0.25 per million tokens on Fable 5.1, a quarter of the Fable 5 rate.
Handle refusals. Configure a fallback. Anthropic lists Opus 4.8 and Opus 5 as the permitted fallback targets.

Where to go from here

If you'd rather start with something lighter, read our guides to using Claude Sonnet 5.5 and prompting Opus 5.5. The Claude prompts page has free prompts for all of them, and the prompt builder fills in the template above.