Which model is worth it is a GET, not a debate

Quarterly planning. Two engineers want the frontier model for every agent. Finance wants the cheap one. Both arguments are made from vibes.

The agent statistics rollup by model ends it. Every agent call through the Gateway is recorded under its container identity, and the rollup shows tokens, sessions, and cost per model across the whole fleet. The frontier model handled 9% of sessions and 61% of spend. The mid-tier model handled the support triage agents at a third of the cost, and nobody complained.

You move triage to the mid-tier model. The servers don't care: they're flat-rate either way. Only the AI Balance notices.

Model budgeting used to be an opinion meeting. Now it's a rollup you read before the meeting.