Fable 5 in Claude Code — what the banner means.

Verbatim, from my own session:

"Complex, long-running work." That's Anthropic's routing table compressed to a single clause — not "for everything," not "your new default." The launch receipts back the phrasing (vendor-curated, so weigh accordingly): Stripe ran a 50-million-line Ruby codebase migration in one day that was scoped at two months by hand. Long horizon, one seat, run until finished. That's the workload class the banner is naming.

"Included in your plan limits until Jun 22." The free-eval clock. Two weeks of frontier capacity inside the subscription you already pay for — which is a private-eval window someone else is funding (Ch 25). The window covers Pro, Max, Team, and seat-based Enterprise; API and consumption-based Enterprise pay per token from day one.

"Then switch to usage credits to continue." The meter. After June 22 the model doesn't leave — the accounting changes. Anthropic says it aims to restore Fable 5 as standard when capacity allows, but you plan on the clause in front of you, not the one promised. The per-token math is on the pricing page: $10/$50 per million, 2× Opus 4.8.

Run /model, pick Fable 5. That's the whole migration inside Claude Code — no API key dance, no settings file. The model id is claude-fable-5; what you get is a 1M-token context window and 128K max output.

If you drive the model through the SDK instead, there's one new sharp edge: an explicit thinking: disabled returns a 400 on Fable 5 even though Opus 4.7/4.8 accept it — omit the param entirely. The full surface, including effort levels and the cache-prefix change, is on the API page. For /model and the rest of the slash-command keyboard, the cheat sheet stays the reference.

You have until June 22 to learn what this model does on your work, free. Don't spend the window on tasks Sonnet already finishes. Spend it where Opus 4.8 makes you wait, retry, or pre-chew the problem:

1. The refactor you've been splitting into chunks for Opus. Hand it the whole thing in one seat and see if the pre-chunking ritual was for the model or for you. SWE-Bench Pro says 80.3% vs Opus 4.8's 69.2% — a signal, not a receipt; your repo is the eval.
2. An overnight run-until-done loop. Ch 38's pattern: /goal with a finish line the transcript can prove, Haiku evaluator deciding "done," a turn-cap tail. The chapter's receipt — a model-bump goal cleared in twelve turns, eighteen minutes, $0.04 of evaluator spend — was the shape with Opus in the main seat. Now put the loop your weaker runs kept failing on the frontier model and check the morning diff.
3. A dynamic-workflows orchestration. The conductor seat holds the plan, the budget, and the merge — that seat earns the bigger model. The fan-out underneath it doesn't (next section).
4. The gnarliest bug-hunt in your backlog. The one with three dead hypotheses pinned to it. Long-horizon investigation is the category the banner names, and FrontierCode Diamond at 29.3% vs 13.4% is the launch table's version of the claim — discount it, then test it.

Ch 29 calls model routing the second-biggest cost lever after caching, and Fable 5 doesn't change the rule — it raises the stakes on it. The split: the main loop, the seat that holds context across the whole task and makes the judgment calls, goes to Fable 5. Fan-out subagents doing bounded work — extract this, verify that, grep those — stay on Sonnet and Haiku. Ch 29's diagnostic was a skill "calling Sonnet 60 times in a row when it should be one Sonnet call orchestrating Haiku tool calls." Same disease one tier up: one Fable 5 conductor orchestrating cheap calls, not ten Fable 5 subagents doing Haiku's job at $10/$50 per million.

And one routing input most operators will misread as a bug: Fable 5 carries classifiers on offensive-cyber, bio/chem, and distillation requests. Trip one and the response comes from Opus 4.8 — Anthropic states users are informed when it happens, and more than 95% of Fable sessions involve no fallback at all. If your security tooling or bio-adjacent sessions keep getting bounced, the product is telling you where that workload lives: on Opus 4.8 directly, at half the rate, with none of Fable's classifiers in the path.

Keep the transcripts from the free window. Every Fable 5 session between June 9 and June 22 is a private-eval data point you didn't pay for — which tasks finished, in how many turns, with how many of your interventions. That archive is what you decide with on June 23, instead of vibes (Ch 25).

The decision itself is Ch 29 math, not sticker math. The price of a model is not the price of a task — if your turn count collapsed on long-horizon work, 2× per token can be net cheaper per finished task. If it didn't collapse on your workload, that's a clean answer too: Opus 4.8 is not deprecated, it stays excellent, and it costs $5/$25.

The banner gave you the routing advice and the deadline in one sentence. The only thing it didn't give you is your own numbers. That part's on you — and until June 22, it's free.