Claude Fable 5 Full Setup Guide

Fable 5 went dark on June 12.

Three days after launch. Gone.

An export control order, a jailbreak Amazon flagged, nobody sure when it was coming back.

It came back July 1.

Same underlying model. Different guardrails wrapped around it.

If you read the first setup guide from June 10, some of that still holds. A chunk of it doesn’t. This is the updated version, built around what’s actually true right now.

Fable 5 and Mythos 5 are still the same model underneath.

Mythos is the raw machine. Fable is the public machine with locks on it.

That part hasn’t moved. What’s new is what the locks do.

Fable now runs a safety classifier aimed at three specific things: offensive cybersecurity techniques, biology and life sciences content, and any attempt to pull out the model’s private reasoning. Trip one of those and Fable doesn’t error out. It quietly falls back to Opus 4.8, finishes your request there, and bills you at Opus rates for that response. Anthropic says this hits under 5% of typical sessions.

Mythos 5, the version with no locks at all, is still restricted to roughly 100 vetted cybersecurity and infrastructure organizations through Project Glasswing. That hasn’t loosened. If you’re reading this from Claude.ai or Claude Code, you’re on Fable, and this whole guide is about that one.

Model ID: claude-fable-5
Context window: 1,000,000 tokens
Max output: 128,000 tokens
Input: $10 per 1M tokens
Output: $50 per 1M tokens
Cached input: $1 per 1M tokens
US-only inference: 1.1x on both input and output

You can reach it through the Claude app, the Claude API, Claude Code, Claude Cowork, Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. It’s also showing up in the model picker for tools like GitHub Copilot and Cursor as they add support.

Don’t assume every one of those surfaces gives you the same limits or the same data terms. Check the actual model picker and the privacy terms before you drop serious work into any of them.