Using Claude Fable 5 in Claude Code: Setup & Best Practices

Claude Code is where Fable 5's long-horizon strengths show up most clearly - multi-hour autonomous runs, self-verification, and far fewer round trips with you. But it is not the default model on any plan, and using it changes both your workflow and your usage math. Here is how to turn it on, what to expect, and how to get the most out of it.

Enabling Fable 5

Inside any Claude Code session, switch models with:

/model fable

That's it. Fable 5 is not the default on any plan - Opus 4.8 (or your configured model) remains the default until you switch explicitly. Fable 5 works with Claude Code's effort settings, so you can pair the model switch with a higher effort level for hard problems (see our effort parameter guide).

What it costs against your plan

In claude.ai subscriptions, Fable 5 counts roughly 2x the usage of Opus toward your plan limits - mirroring its 2x API price. Two scheduling facts matter right now:

Free until June 22, 2026: Fable 5 is included at no extra cost in Pro, Max, Team, and Enterprise plans during the launch window, per Anthropic.
After June 22: it requires usage credits billed at API rates, until Anthropic restores it as a standard part of subscription plans "when capacity allows."

Practical takeaway: the next two weeks are a free trial on your real codebase. Run your hardest backlog items through /model fable now and decide with data whether the 2x multiplier is worth it for your daily work.

What actually changes

Anthropic's pitch is that Fable 5 "tackles your biggest challenges with fewer check-ins needed," and early usage bears that out. The practical differences from Opus 4.8 in Claude Code:

Longer autonomous stretches. Fable 5 sustains multi-hour - in some cases multi-day - agent sessions, planning across stages and delegating to sub-agents rather than stalling for guidance.
Self-verification. It tests its own work before reporting back. Salesforce's launch quote captures it: "at the highest effort, reflects on and validates its own work."
Fewer turns per result. Amazon reported "more capable engineering in fewer turns" - which also softens the 2x usage multiplier in practice.
Better intent reading. Replit's CTO: it "understands what builders mean, not just what they type... one-shots."

Ethan Mollick's early experiments are the most vivid demonstration: 12-hour unattended runs against a written spec, and complete, playable video games from a single prompt - a strategy game called "Strata" and "Duino," a game built from a Rilke poem. His verdict: Fable 5 "outperformed basically every other public model I have used by a considerable margin."

Best practices for long-horizon sessions

Front-load the entire task spec. Fable 5 rewards a complete brief in the first turn - requirements, constraints, definition of done - far more than iterative drip-feeding. Mollick's 12-hour runs started from a full written spec, not a conversation.
Use high or xhigh effort for hard work. Adaptive thinking is always on, but effort controls how deep it goes. Routine edits run fine at lower effort; architecture, migrations, and gnarly debugging justify high or xhigh.
Give it output headroom. Fable 5 supports up to 128K output tokens. Long autonomous runs produce long transcripts of plans, diffs, and test output - don't strangle them with tight max_tokens limits (and remember 128K output requires streaming on the API).
State outcomes, not steps. Phrase goals the way you'd phrase a /goal: "the test suite passes and the API handles auth via OAuth" beats a numbered list of micro-instructions. Fable 5 plans well; let it.
Let it verify itself. Tell it explicitly to run the tests, exercise the feature, and fix what it finds before reporting. Its self-validation is a genuine capability - invoke it.

When to stay on Opus 4.8

Fable 5 is not a universal upgrade for Claude Code work. Stay on Opus 4.8 (or Sonnet) when:

The work is routine and high-volume - small fixes, boilerplate, mechanical refactors - where Opus already one-shots reliably and the 2x usage cost buys nothing.
You're near your plan limits and need the rest of the billing cycle to stretch.
You want interactive, short-leash pairing rather than handing off a large task; the gap between models narrows when you're steering every turn anyway.

The safety-flag model switch

One setting specific to the Fable era: "Switch models when a message is flagged." Fable 5 ships with real-time safety classifiers, and when one flags a message, this toggle automatically switches you to a different model so the conversation can continue rather than dead-ending. With the toggle off, a flagged message simply stops.

The trade-off is transparency. The fallback fires in under 5% of sessions by Anthropic's count, but critics - most prominently Nathan Lambert - object to a model that silently gets less capable mid-conversation. If you enable the switch, watch for sudden changes in output quality on security- or biology-adjacent work; if you'd rather see the refusal explicitly, leave it off. Our safety deep dive covers the classifier system and the false-positive debate in full.