claude-fable-5 — the API page.

From Opus 4.8 or 4.7: the request surface is identical. Swap the model string, ship. Adaptive thinking carries over, the effort ladder carries over, your tool schemas carry over. The one exception is the new 400 below — and it only bites code that was explicitly disabling thinking.

From 4.6 or older: the 4.7-era removals apply to you all at once. Three things in your old request will now return a 400:

• budget_tokens — gone. Adaptive thinking replaced manual thinking budgets; send thinking: {type: "adaptive"} and pick an effort level instead.
• temperature, top_p, top_k — removed. There is no sampling knob to tune on this surface.
• Last-assistant-turn prefills — rejected. If your prompt scaffold ends with a primed assistant turn, restructure it into the system prompt or the last user turn.

None of these are Fable-specific — they're the Opus 4.7/4.8 surface, and Fable 5 inherits it unchanged. The head-to-head on whether the swap is worth 2× lives at Fable 5 vs Opus 4.8.

Here is the single place where Fable 5's API diverges from Opus 4.8: an explicit thinking: {type: "disabled"} returns a 400. Opus 4.7 and 4.8 accept it. Fable 5 rejects it. If you don't want to send a thinking config, omit the parameter entirely — don't send disabled.

# before — Opus 4.8 (this exact request 400s on Fable 5)
resp = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=8192,
    thinking={"type": "disabled"},   # accepted on 4.7 / 4.8
    messages=[...],
)

# after — Fable 5
resp = client.messages.create(
    model="claude-fable-5",
    max_tokens=128000,                   # the Fable 5 output ceiling
    thinking={"type": "adaptive"},       # or omit the param entirely
    output_config={"effort": "xhigh"},   # for agentic work
    messages=[...],
)

Effort is the knob that survives: low / medium / high / xhigh / max, all supported. For long-horizon agentic work, xhigh is where the launch table's headline number lives — FrontierCode Diamond was reported at xhigh. For bulk extraction, drop the effort, not the model's thinking.

The quietest spec change is the one that touches your bill. The minimum cacheable prefix on Fable 5 is 2,048 tokens — on Opus 4.8 it's 4,096. A 3K-token system prompt with a cache_control breakpoint caches on Fable 5 and silently doesn't on Opus 4.8. No error either way. The invoice is the only witness.

At $10 per million input tokens, this floor matters more than it did at $5. Cache reads run at ~0.1× — so a cached prefix that re-fires across an agent loop is the difference between paying frontier price once and paying it every turn. If your prompts hover in the 2K–4K band, Fable 5 is the first Opus-class surface where they cache at all. The full token math is Ch 29's territory.

Fable 5 ships with classifiers in front of three areas — offensive cyber, biology/chemistry, capability distillation. On the Messages API the default behavior is clean: a blocked request returns an error, not a model response — and you aren't charged for it. Three ways to handle it:

• Opt in to fallback on the Messages API — blocked requests continue on Claude Opus 4.8, billed at Opus pricing ($5/$25), not Fable pricing.
• Roll your own — catch the error and re-route with the Anthropic SDK's support for exactly this.
• Managed Agents — fallback is built in; nothing to wire.

Per Anthropic, more than 95% of Fable sessions involve no fallback at all. The operator read for API builders: this is routing information, not fine print. If your product is security tooling or bio-adjacent, expect Opus-4.8-grade answers on those paths by design — the starred rows in the launch table already priced this in. Route those workloads to Opus 4.8 directly at $5/$25 instead of eating a classifier round-trip to end up there anyway. The decision table is on the vs-Opus page.

Two more lines of fine print that gate real deployments: Fable 5 requires a limited 30-day retention period — retained data is used only to detect and prevent serious misuse, never to train Claude — and admins must accept updated terms in the Claude Console before the model works. Budget five minutes for the Console before you budget an afternoon for the migration.

The launch detail most API builders will skim past is the one with the best economics: Fable 5 is available as an advisor model. Faster, lower-cost worker models call it mid-task to check their plan and evaluate their work — Anthropic's own framing, and their own claim is that it leads to improved performance. You don't put the $10/$50 model in the driver's seat; you put it at the judgment gate.

If you've read this book, you've seen this shape before. It's the conductor-and-judge split from Ch 6 and the second-opinion loop from the swarm patterns — workers fan out on Sonnet-tier, one expensive model verifies plans and grades output. The difference is that it's now a first-party API primitive instead of an orchestration pattern you build yourself. Same logic as the Amdahl gate in the research notes: the bottleneck is judgment, so spend the frontier tokens on judgment.

And the pairing Anthropic built for the long-horizon story: Claude Managed Agents, their harness for long-running agentic work, is now in public beta — Fable 5 works with it out of the gate, no changes needed. Fallback included.

Ch 30 wrote this exact scenario before the model existed: a 34-line SDK-direct file where the next frontier model is a one-line swap — the cache control still works, the tool schema still works, the retry config still works. That chapter used a hypothetical id. Today the hypothetical has a real name: claude-fable-5.

The other half of the prediction held too. SDK-direct paths can test Fable 5 today. Framework-shaped paths — CrewAI, LangGraph, the agent runtimes — wait for the framework to publish support: provider config, adapter updates, retry semantics for the new 400. That's structural lag, not Fable-specific lag, and it's the same lag Ch 30 flagged at the Mythos disclosure. Ch 2 makes the stack fixed lanes; Ch 30 makes the model inside each lane a variable — and this morning is what that discipline was for.

If the model id is hardcoded in nine places across your stack, today is the day you find out. Fix that first, swap second, and let the tier list decide which workloads earn the 2× rate. The 400 will find you if you earned it.