Someone routed a classification step to Opus. The cost spread is 200x.

A copy-paste from a benchmarking experiment, an env var pointing at the wrong default, a prompt with "use the most capable model" language landing on an autorouter. The result is the same: a step that should run on Haiku quietly bills Opus rates. Calcis shows the cost spread in the PR diff before the regression ships, with a one-line revert hint when the new model is overkill for the workload's complexity. Longer write-up plus a CI recipe landing here next.