First time seeing savings

Coder measures each visible user turn. A visible prompt is counted once even if the client makes several model calls.

What you should see

• Input tokens actually sent
• Cache and output where the dashboard provides those tabs
• A counterfactual only on the first eligible invocation of a conversation

What you should not expect

There is no cross-user median published on this site. The only approved numeric result is a four-turn controlled E2E measurement documented on Telemetry and savings. Results vary by repository, client, model, and task.

Local model use is not converted into a dollar saving when no paid model cost applies.