What Fable 5 actually is
Fable 5 (claude-fable-5, codenamed Mythos) is Anthropic's most capable model — a new tier that sits above Opus. Same 1M-token context as Opus 4.8, same 128K max output, but it's tuned for the hardest work: deep research, complex judgment, and large multi-step projects.
- Model id —
claude-fable-5 - Context / output — 1M tokens in · up to 128K out (stream for big outputs)
- Price — $10 / 1M input · $50 / 1M output — 2× Opus 4.8
Is it actually better? (and the catch)
Early hands-on reviews put it in a different class for senior-level engineering. In Every's week-long test it scored 91/100 on their senior-engineer benchmark — vs 63 for Opus 4.8 and 62 for GPT-5.5.
- Reach for it: deep research, architecture, large refactors, tricky judgment calls.
- Skip it: chat, classification, simple edits, high-volume jobs — cheaper models win on speed + cost.
The model picker — which Claude for the job
You rarely want one model for everything. Match the model to the task and you cut cost without losing quality.
| Model | Best for | Context | $/1M in·out |
|---|---|---|---|
| Fable 5 | Hardest research, judgment, big projects | 1M | $10 · $50 |
| Opus 4.8 | Strong default for agentic + coding | 1M | $5 · $25 |
| Sonnet 4.6 | Best speed/intelligence balance | 1M | $3 · $15 |
| Haiku 4.5 | Fast, cheap, simple tasks | 200K | $1 · $5 |
Switch to Fable 5 without getting a 400
Fable 5 shares Opus 4.8's request surface, with a few parameters removed. Update these and you're done:
- Set the model string to
claude-fable-5. - Use adaptive thinking:
thinking: {type: "adaptive"}. The oldbudget_tokensis gone (returns 400). - Delete
temperature,top_p,top_k— all removed. Steer with the prompt instead. - Fable-5 quirk: do not send
thinking: {type: "disabled"}(400). To run without thinking, just omit thethinkingfield entirely. - Tune
output_config: {effort: ...}—low→max(incl.xhigh). Start athigh; reservemaxfor the hardest jobs. - Streaming for large outputs (>~16K tokens) to dodge HTTP timeouts.
output_config.format for structured output). Cacheable-prefix minimum is 2048 tokens.5 ways to put it to work today
- Research analyst — Dump 10+ sources in the 1M window; ask for a cited, decision-ready brief.
- Architecture partner — Hand it the whole task spec up front; let it plan the build at high effort.
- Large refactor — Point it at a big codebase and a clear goal — it holds long-horizon context well.
- Judgment calls — Use it where being right matters more than being cheap (contracts, tradeoffs, reviews).
- Route, don't default — Send only the hard 10% to Fable 5; everything else to Sonnet/Haiku.
Get the next drop
New AI build guides + the occasional bonus template. No spam, unsubscribe anytime.