Model routing
Updated 2026-07-11
Questions this answers
- My AI costs are climbing faster than the value I'm getting, what do I cut first?
- Do I really need the most expensive model for every task?
- How do teams keep AI spend under control without slowing down?
- Is there a low-effort way to reduce token costs?
The Fix
The simplest cost lever that actually works is model routing: send each task to the cheapest model that can still do it, instead of defaulting to the top model for everything. Plenty of work, drafting, summarizing, simple edits, doesn't need your most expensive model.
The Pragmatic Engineer reported a mid-size software company that cut costs around 30% just by changing its default model to a cheaper one, with no loss of momentum. In Claude Code you can set this with /model, and tools like OpenRouter let you route across providers.
Fit it inside a simple frame: spend, then measure, then adjust. Experiment freely in the short term, check your outcomes monthly, and when spend rises without matching results, close the gap. Reserve the big model for the tasks that genuinely need it, like the review step in a critique loop.
When to Use It
Look at routing when your AI bill is growing faster than the value, or before you roll a workflow out to a whole team where costs multiply. Don't over-optimize a hobby project where the spend is trivial. And keep an eye on quality: route down until output slips, then settle one step back up.
Best Practices
Pragmatic Engineer: token spend breaks budgets
newsletter.pragmaticengineer.com
The write-up on how changing the default model cut costs ~30%, plus the spend-measure-adjust framing.
OpenRouter
openrouter.ai
A single interface for routing tasks across many models and providers so you can pick by cost and fit.