Model Routing: The Right Model for the Right Task

Not every task needs the strongest (and most expensive) model — model routing saves cost without sacrificing quality.

The problem

A strong model like Claude Opus is great at complex reasoning, but expensive and often slower. A cheap, fast model like Claude Haiku is often just as good for simple, mechanical tasks — at a fraction of the cost. Using the most expensive model for everything burns money without buying any quality.

What model routing means

Model routing means sending different tasks to models of different strength: a strong model takes the orchestrator role — it plans, breaks down complex tasks, and makes hard decisions. Cheaper models handle mechanical subtasks: classifying, formatting, simple summaries, routine code.

Anthropic's own recommendation

Anthropic describes routing as its own agent pattern: route easy or common requests to smaller, cost-efficient models like Claude Haiku, and hard or unusual requests to more capable models like Claude Sonnet. This works well when tasks fall into clearly distinguishable categories and that classification can be made reliably and automatically.

The cost lever, concretely

Price differences between models are substantial: a cheap model can cost a fraction, per million tokens, of what the strongest available model costs. Across thousands of requests a day, that difference adds up to a large gap in the bill — without the simple tasks getting noticeably worse.

EXAMPLE

A coding agent gets the task 'clean up every console.log call in the project and write a summary of which files are complex enough to need manual review'. A cheap model mechanically removes the console.log lines across hundreds of files. A strong model (the orchestrator) then evaluates only the few genuinely unclear cases and writes the summary.

🛠️ EXERCISE — TRY IT YOURSELF

Analyze one of your own (or a hypothetical) multi-step workflows and decide which subtasks a cheap model could take over.

List every step of your workflow (e.g. research, classification, formatting, final decision).
Mark for each step: does it need real reasoning/creativity, or is it mechanical/unambiguous?
Roughly estimate the cost savings if the mechanical steps move to a cheaper model (use the pricing table from the sources).

✅ SELF-CHECK

☐ Is there at least one step that's clearly mechanical and currently running unnecessarily expensive?
☐ Is the 'easy vs. hard' classification itself reliable enough to automate?
☐ What happens if the cheap model fails on a hard case — is there a fallback?

QUICK QUIZ

According to Anthropic's own description, when does model routing pay off in particular?