model-routing

name	model-routing
description	Pick the right model per call, not per project — route Opus/Sonnet/Haiku, GPT-5/4o/mini, Gemini Pro/Flash by task, and cut cost without losing quality. Use when the user is choosing model tiers, building a router, or debating Opus-only vs mixed-tier deployments and mentions model selection, model router, cascade, fallback, cheap-first, draft-then-verify, or asks "which model should I use?" / "do I need Opus for this?".
tags	["cost","latency","architecture","model-selection"]

Model Routing

Pick the model per call, not per project.

A single-tier deployment is either overpaying for easy turns or underdelivering on hard ones. The lever is matching model capability to the specific decision the call is making — and only escalating when the cheaper model demonstrably fails.

When to use this skill

• The user is choosing a model for a new agent.
• The user is on the most expensive tier for everything and the bill is climbing.
• The user is on the cheapest tier and quality is borderline.
• The user is building or evaluating a router / cascade / draft-then-verify pipeline.
• The user mentions Opus, Sonnet, Haiku, GPT-5, GPT-4o-mini, Gemini Pro/Flash, model fallback, or hybrid inference.

Decision flow

1. Is the task single-shot, easily-verified, and bounded? → cheapest tier. Don't romanticise quality where it doesn't move the needle.
2. Is the task agentic (multi-step tool use, planning)? → the planner needs the strongest model; the executor can be smaller.
3. Is the task latency-critical (TTFT-sensitive UX)? → cheapest fast model that clears the quality bar. Cascade if quality wobbles.