Claude Fable 5: The Field Guide

Fable 5 (claude-fable-5, the public release of the Mythos line) is Anthropic's most capable widely-released model. Confirm these before you wire anything up — the rest of this guide builds on them.

• Model id — claude-fable-5 — there is also claude-mythos-5, the same model without the safety classifiers, but it's invite-only (Project Glasswing). For everyone else it's claude-fable-5.
• Context / output — 1M tokens in, 128K out. Stream anything over ~16K output or the SDK times out the request.
• Price — $10 / 1M input · $50 / 1M output — 2× Opus 4.8's $5/$25, 3.3× Sonnet's $3/$15, 10× Haiku's $1/$5.
• Quiet cost multiplier — Fable 5 uses the Opus-4.7 tokenizer: the same text is ~30% more tokens than on older models. So your effective bill is closer to ~2.6× Opus on identical inputs, not 2×. Re-measure with count_tokens before you trust an old budget.

The honest answer is rarely as your default. Match the task to the cheapest model that clears the quality bar — Fable 5 earns its premium on a narrow set of jobs. Use this as your routing rule:

Same base_url, same SDK — you switch models by changing one string. Here's what each rung costs and what it's for.

Fable 5 shares Opus 4.8's request shape but rejects a few parameters Opus tolerated. Each one below is a hard 400 if you leave it in. Steps 1–4 are the ones that bite during migration.

1. Set the model to claude-fable-5.
2. Delete temperature, top_p, and top_k. All three are removed — any of them returns a 400. Steer with the prompt instead.
3. Remove thinking: {type: "enabled", budget_tokens: N}. Budgets are gone; sending one is a 400. Thinking is always on (adaptive) — just omit the thinking field, or set thinking: {type: "adaptive"}.
4. Do NOT send thinking: {type: "disabled"}. This one is Fable-specific: it's accepted on Opus 4.8/4.7 but a 400 on Fable 5. There is no "thinking off" — control depth with effort instead.
5. Set depth via output_config: {effort: ...} — low · medium · high (default) · xhigh · max. Start at high; only go xhigh for the most capability-sensitive work, max for genuinely frontier problems (it can overthink).
6. Drop any last-assistant-turn prefill — also a 400. Use output_config.format (structured output) or a system-prompt instruction to shape output.
7. Stream for outputs over ~16K tokens so you don't hit an HTTP timeout, and give max_tokens real headroom (it caps thinking + text combined).

Unlike Opus, Fable 5 runs safety classifiers that can decline a request — and a decline is not an error. It comes back as a successful HTTP 200 with stop_reason: "refusal" and an empty (or partial) content array. Any code that reads response.content[0].text without checking stop_reason first will crash on a refusal.

• What triggers it — Offensive-security/exploit work (cyber), bio/lab methods (bio), help building competing models (frontier_llm), or asking it to dump its own reasoning as text (reasoning_extraction). Benign security and life-sciences work can trip these too.
• How to detect it — Branch on stop_reason == "refusal" — NOT on content. stop_details.category names the policy, but it can be null, so don't key your logic off it.
• Billing quirk — A refusal before any output costs nothing and doesn't count against rate limits. A mid-stream refusal bills the input + already-streamed output — discard the partial.
• Monitoring trap — Refusals are 200s, so dashboards built on error/5xx rates never see them. Emit your own metric per refusal.

Both of these produce confusing failures that have nothing to do with your prompt.

• Every request 400s out of nowhere — Fable 5 requires 30-day data retention and is not available under zero-data-retention (ZDR). If your org is on ZDR — or any retention below 30 days — every Fable 5 call returns 400 invalid_request_error, even a perfectly valid one. Check the org's retention setting before you debug the payload.
• Your cost / token math is suddenly off — The ~30% tokenizer inflation means token counts, context budgets, and max_tokens values measured on Opus/Sonnet/Haiku don't carry over. A prompt that fit comfortably before can blow your budget. Re-run count_tokens passing model: "claude-fable-5" — the response reports counts under both the new and old tokenizers so you can see the delta.
• Migrated prompts feel worse, not better — Over-prescriptive scaffolding written for older models can degrade Fable 5. It follows instructions tightly and plans well on its own — strip the step-by-step hand-holding and re-test; a short "act when you have enough info; don't refactor beyond the task" beats a long checklist.

Concrete patterns that play to Fable 5's strengths (long-horizon, autonomous, self-verifying) while keeping the bill sane.

• Research analyst — Load 10+ sources into the 1M window at high effort; ask for a cited, decision-ready brief. This is the textbook Fable 5 job.
• Architecture partner — Give the WHOLE spec up front in one well-specified turn, then let it plan and build. It rewards a clear goal more than mid-task nudging.
• Long autonomous run — Plan for minutes-long (sometimes hours-long) turns: stream, show progress, and check in asynchronously instead of blocking. Add "audit each progress claim against a tool result before reporting it" to kill fabricated status updates.
• Parallel sub-agents — It dispatches sub-agents reliably — delegate independent subtasks and keep working rather than spawning-and-blocking.
• Route, don't default — Send the hard ~10% here; everything else to Opus 4.8 / Sonnet / Haiku. One endpoint, one-line model swap (see below).

The whole point of the model-picker table is that the model is a config value you swap per task. To make that real you need one place to call all of them — and if you're serving customers, a way to meter and bill what they use. Knotie's AI Gateway is exactly that: an OpenAI-compatible endpoint. Point the standard OpenAI SDK at https://api.knotie.ai, change the model name, and you reach budget, mid, and premium tiers (Claude — including Fable 5 — plus GPT and Gemini families) without a new provider integration per model.

• One-line model switch — Swapping claude-fable-5 for claude-opus-4-8 (or a GPT/Gemini model) is the model name + base_url — not a new SDK, not new auth.
• Per-customer metered keys — Mint virtual keys, restrict each key to specific models from a tiered list, set a profit markup, and bill usage on credits — under your own brand and domain.
• Guardrails before a demo — Restrict a key to mid-tier only, add domain whitelisting, and watch spend per key — so a client demo can't quietly run the $50/1M model on everything.