Published guides claim that routing work between the largest AI models and smaller ones cuts inference cost by 40% to 60%. This review tested four of those strategies against 30 days of session transcripts and the code of the scheduled jobs before adopting any of them.
- The four strategies were a router that classifies each request, a large model directing smaller workers, a cheap first pass that escalates on a weak result, and a tighter harness around a small model. The savings figures in the two guides cited no study. The one measured benchmark found showed a 74% cost reduction alongside an accuracy loss of about six points.
- The transcripts covered 10.2 billion tokens. The main conversation accounted for 96% of them and delegated helpers for 4%, so moving every read-only helper to a smaller model would have saved between 2% and 4%. Cost follows what enters the main conversation and how long the session runs, because each later step reads all of it again.
- Two changes followed. Every session started at about 99,000 tokens before any work, so a duplicate tool connector was disabled in the configuration of the project that never uses it. A status check that loaded a large document into the main conversation at the start of each session was rewritten to run in a read-only helper that returns only the differences.
- The scheduled jobs already ran on mid and small tier models, each fixed in its own code, so a router would have added a component without changing a choice. Escalation logic stays in each job beside the validator that defines a weak result. A new monthly check compares all 13 pinned models with the published model list, prices, and retirement dates, and opens one task for a person when something changes.