haiku-4-5

Claude Haiku 4.5 (claude-haiku-4-5) protocol — fast/cheap fan-out patterns, what to route to it and what not, thinking config. Trigger when working with, routing to, or writing code that calls Haiku 4

SKILL.md

GitHub

Haiku 4.5 — Speed-and-Volume Protocol

Model ID: claude-haiku-4-5 (full: claude-haiku-4-5-20251001) · 200K context · 64K max output · $1 in / $5 out per MTok. The fastest, cheapest current model — its job is volume, latency, and fan-out, not depth.

Staleness guard

Model facts here (ID, pricing, limits, API rules) were verified 2026-07. Haiku 5.5 (haiku-5-5 skill) superseded this model on 2026-10-07; use this file only for projects pinned to 4.5. If the API contradicts this file: verify against current docs (Models API / platform.claude.com or the claude-api skill), follow the live source, and log the correction in learnings.md.

Self-improvement (do this first and last)

  1. At start: read learnings.md in this skill's folder if it exists. Apply relevant lessons.
  2. At end of every use: append one dated bullet — what worked, what wasted tokens, what you'd prompt differently. Merge instead of duplicating; delete disproven bullets.

Hard API rules (older-generation surface — different from Opus/Sonnet 5)

  • Thinking uses the legacy config: thinking: {type: "enabled", budget_tokens: N} — N ≥ 1024 and strictly less than max_tokens. Adaptive thinking is not available; omit thinking for no thinking (the usual choice for Haiku work).
  • effort errors on Haiku 4.5 — don't send output_config.effort.
  • temperature/top_p allowed (not both), unlike Opus 4.7+ / Sonnet 5.
  • 200K context and 64K max output — the smallest limits in the lineup; chunk accordingly. Stream above ~16K output.
  • Separate rate-limit pool from older Haiku models.

Route TO Haiku

  • Classification, extraction, tagging, routing, dedup, summarizing individual documents.
  • Subagent fan-out: file sweeps, grep-and-report, per-item checks across many items — dozens of parallel Haiku calls cost less than one Opus turn.
  • Latency-critical UI paths (autocomplete, quick answers, validators).
  • High-volume batch pipelines (pair with the Batches API for another 50% off).

Do NOT route to Haiku

  • Multi-step architecture or debugging, long-horizon agent loops, anything where a wrong answer costs more than the Opus/Sonnet premium. Retrying Haiku three times on hard work costs more than one Sonnet call and still fails.

Prompting patterns that matter on Haiku

  • Be maximally explicit. Small models reward rigid structure: enumerate steps, give the exact output format, include one or two worked examples. What Fable treats as over-prescriptive is exactly right here.
  • Constrain output: use structured outputs (output_config.format json_schema) or enum-field tools for classification — don't parse free text.
  • Tight max_tokens (e.g. 256 for classification) — Haiku's cost is per token like everyone else; short outputs are the point.
  • One task per call. Compound instructions degrade fast — split them.

Token discipline

  • Cache the shared prefix across the fan-out (min cacheable prefix on Haiku 4.5 is 4096 tokens).
  • Send only the slice of context each call needs — never the whole file set to every worker.
  • Aggregate Haiku outputs with a stronger model rather than asking Haiku to synthesize across items.