Anthropic shipped Claude Fable 5.1 today.

It is their public model for long coding jobs and knowledge work.

There is also Claude Mythos 5.1. Same model. Different safeguards. Mythos is only for vetted US orgs in cyber and life sciences. You almost certainly want Fable.

Model id: claude-fable-5-1

1) What actually changed

Fable 5.1 is not a new chat toy.

It is the follow-up to Fable 5 for work that lasts longer than one sitting.

Official numbers that matter if you write code:

Terminal-Bench 4.0: 55.8% vs 42.0% on Fable 5

CursorBench 3.2: 73.4% vs 70.5%

Cache reads: $0.25 per million tokens (was $1 on Fable 5)

Headline API price did not move.

$10 per million input. $50 per million output.

The cheaper cache is the whole cost story.

Typical token-billed work is about 25% less than Fable 5.

Heavy agent loops can land around 45% less, because those jobs reread the same context over and over.

It defaults to High effort in Claude Code, and Medium on claude.ai and Cowork.

Effort is how much it thinks per step. You can set low, medium, high, xhigh, or max.

2) When to reach for it

Use Fable for the jobs you would normally slice into a bunch of smaller prompts.

Root-cause a crash. Hold a refactor across a fat repo. Keep an investigation going without losing the thread.

Claude's own docs say to describe the outcome, not the steps, and to skip the "please test it" reminders. It already verifies more than the smaller models.

Do not make it your daily Sonnet replacement for a 12-line fix. It is not the default on any plan. You have to pick it.

3) Claude Code path

You need Claude Code v2.1.255 or later. Run claude update if the picker does not show it.

Then type:

/model fable

Or start a session with:

claude --model fable

The fable alias now resolves to 5.1. Before v2.1.255 it pointed at Fable 5.

If you previously saved claude-fable-5 in user settings, v2.1.255+ will auto-update that alias the first time you launch. Project and managed settings stay pinned.

If the picker says "Requires usage credits", that is real.

Fable can bill extra depending on plan and seat.

Interactive sessions should ask before they spend.

Headless -p and the Agent SDK will not ask. They just bill.

Thinking cannot be turned off on Fable 5.1. Effort is the control. Not a thinking toggle.

Context is 1M tokens on the Anthropic API by default.

4) API / cloud path

On the Claude API the id is:

claude-fable-5-1

It is also on AWS, Google Cloud, and Azure today.

On claude.ai it is rolling to Pro, Max, Team, and Enterprise.

If you pin models in Bedrock or Vertex, pin the 5.1 id on purpose. A leftover Fable 5 pin will not move you.

Prompt caching is why the bill dropped. Do not disable Fable prompt caching unless you have a reason. That is how you throw away the 25-45% save.

5) Fable vs Mythos, and the classifiers

Fable and Mythos are the same weights.

Fable is the one with the public safeguards. Mythos loosens cyber and biology restrictions for people Anthropic has actually vetted.

On Fable, biology-flagged requests can fall back to Opus 5. Cyber-flagged requests can fall back to Opus 4.8. That is expected routing, not your account getting banned.

Anthropic also opened a narrow door: Fable 5.1 can help find software vulnerabilities. It still will not help you write exploits. They claim about 60% fewer cyber false positives per Claude Code session vs the old Fable 5 classifiers.

If you do offensive security or research biology all day, Fable will keep bouncing you. That is what Mythos's access programs are for.

6) A sane first job

Do not "try Fable on everything."

Pick one job you already do that dies when the model loses the plot.

Examples that fit:

A flaky test you have not been able to pin down

A refactor that touches more than one package

"Read this repo and tell me why prod only dies on Tuesdays"

Update Claude Code. Switch with /model fable. Give it the outcome. Let it run.

If the session is a 40-second autocomplete, switch back to Sonnet.

That's Fable 5.1.

A long-job model with a cheaper cache, not a new default.

Use it when the work is big enough that losing the thread is the actual problem.

Official writeup: anthropic.com/claude-fable-and-mythos-5-1