Cheap-then-deep routing: Haiku 4.5 first, Opus 4.7 only when it earns it

Two-tier routing. Cheap model triages. Expensive model only on escalation. Most traffic is easy — stop paying Opus rates for it.

Run this as a gate before any expensive call:

You are a triage router. Read the user request below.
Answer ONLY with a JSON object: {"confidence": 0-1, "answer": string|null}.

Rules:
- If you can fully answer with high confidence, set answer to your response and confidence >= 0.8.
- If the request is ambiguous, multi-step, or needs deep reasoning, set answer to null and confidence < 0.5.
- Never guess. Low confidence is cheaper than a wrong answer.

Request:
{{user_input}}

Wire it: if confidence >= 0.8, ship answer. Else replay the original request to Opus 4.7. One branch. No orchestration framework, no router model zoo.

Receipt

• Triage model: Haiku 4.5 (claude-haiku-4-5)
• Escalation model: Opus 4.7 (claude-opus-4-7)
• Triage cost: ~320 in + 40 out tokens ≈ $0.0004
• Blended cost across mixed traffic: ~$0.004 / call (≈70% cheaper than all-Opus)

Why it works

• Easy traffic dominates. 80% of calls never need a frontier model — Haiku closes them for a tenth of a cent.
• Confidence is self-reported but calibrated. Forcing a 0–1 score makes the model commit instead of bluffing.
• One escalation hop. No embeddings, no classifier to retrain, no second system to keep in sync.

Failure mode

• Overconfident Haiku. It occasionally rates a wrong answer 0.9. Sample 1% of shipped answers and grade them; if the false-confident rate climbs, raise the threshold to 0.9.
• Escalation storms. A bad input class can push everything to Opus. Alert when the escalation rate crosses 35%.

Cost to test: $0.004 / call.