Model Routing & Cascades
Routing moves each request to the cheapest model that clears the quality bar. Mature 2026 programs route 60–90% of traffic away from their flagship model with no measurable quality loss.
1. The three routing strategies
Static routing (start here). Map endpoint → model by task class. No runtime classifier, no latency cost, easy to reason about. Captures most of the available savings.
Cascade routing. Try the cheap model; escalate on a confidence or validation failure.
expected_cost = c_small + p_escalate × (c_large + overhead)
A cascade only wins when p_escalate is low. With a small model at 1/20th the price, escalation must stay under roughly 50% for the cascade to beat always-large; under 20% for a comfortable win.
Learned routing. A small classifier predicts difficulty and picks a tier. Worth it above roughly 1M requests/month; below that, the maintenance and misroute risk exceeds the savings.
2. Task-to-tier mapping
| Task class | Tier | Notes |
|---|---|---|
| Classification, routing, tagging, PII redaction | Nano / small | Often a fine-tuned or open-weight 3–8B model |
| Extraction, summarisation, format conversion | Small | Strict schema; low reasoning |
| Drafting, rewriting, standard code edits | Mid | Cache the style guide prefix |
| Multi-file refactor, hard analysis, planning | Flagship | Reasoning budget medium/high |
| Adversarial, safety-critical, legal review | Flagship + human | Never auto-approve |
3. Escalation signals that actually work
- Schema validation failure (deterministic, free).
- Self-reported confidence below threshold — weak alone, useful combined.
- Judge-model disagreement on a sampled subset, not on every call.
- Downstream tool failure or empty retrieval.
- User-visible correction or regenerate click — the strongest label you own.
Avoid escalating on response length or on the model saying "I'm not sure" in prose.
4. Fallback vs routing
Keep them separate in code and in metrics:
- Routing is a cost decision, made before the call.
- Fallback is an availability decision, made after a 429/5xx.
Fallback should prefer a different provider at similar quality, not a cheaper tier — silently degrading quality during an incident is how routing programs lose trust.
5. Quality gates before a swap
- Golden set of 200+ graded real requests per endpoint.
- Head-to-head with the incumbent; require parity within a defined tolerance.
- Shadow traffic for 3–7 days, comparing cost, latency p95, retry rate.
- Canary at 5% → 25% → 100% with automatic rollback on retry-rate regression.
- Record the decision in an ADR with the measured numbers.
6. Common failure modes
- Routing on prompt length rather than task difficulty.
- Letting the classifier itself cost more than the savings (use a tiny model or heuristics).
- Cascades that re-send the full context on escalation instead of reusing a cached prefix.
- No per-tier dashboards — you cannot defend routing without per-tier quality metrics.