Model Router: Claude Picks the Right Model for Every Step, Automatically 🚦

Most people run every step of a task on the same model. Planning, reading files, pulling data, formatting a report, debugging HTML, all of it on the most expensive tier. Sessions also run for days, and every new turn re-reads the whole history. The most expensive failure is still a different one: a perfect execution of the wrong task.

Before any tool call, the skill restates what it understood in at most 8 lines: goal, deliverable, scope, riskiest assumption, a tiered plan and the effort per phase. You confirm once, then it works without further check-ins. Most cost tools tell you to switch models and leave the switching to you. Model Router does the routing itself: it never asks which model to use. Planning and final synthesis stay on the main model. Execution goes one tier down (Sonnet from Opus or Fable, Haiku from Sonnet), mechanical work goes to Haiku from any tier. Multi-source gathering fans out in parallel with a "conclusions only, no raw dumps" contract. It recommends the right effort level at each phase change, keeps large data pulls out of the main context, and stops at about 25 turns with a handoff file so the next session starts clean. Every task ends with a one-line routing summary.

Premium reasoning where it matters, volume pricing everywhere else. Haiku output costs roughly 1/20 of Sonnet, so moving extraction, summaries and lookups down a tier adds up fast. The understanding check removes the costliest waste of all, which is redoing work that answered the wrong question.

Full routing needs a harness that can spawn subagents with a per-call model choice, for example Claude Code, Cowork or the Claude Agent SDK. All model names and prices sit in one registry table. When a model ships or is retired you edit only that table. Overrides are plain phrases: "just do it" skips the understanding check, "session stop at N turns" changes the stop.

Routing work to cheaper subagents is automatic. The skill cannot change your main session's model, the effort setting or a scheduled task's model field, so for those it gives a one-line recommendation at the right moment. Without subagents the delegation steps are skipped, while the understanding check, effort advice and session discipline still apply. Prices and benchmarks were verified on 13 Oct 2026 and change with every release, so check the registry against Anthropic's pricing page before quoting numbers.