3.3.1. Cost Management: Budgets, Quotas, and Token Economics
💡 First Principle: AI cost is dominated by tokens, and the controls that actually bound spend operate at different points — budgets/alerts tell you after the fact, TPM ceilings stop runaway loops in real time, tagging attributes cost, and model right-sizing cuts the per-token rate at the source. A complete answer usually combines them; a single control rarely suffices.
The four most effective spend controls: configure budget alerts (e.g., 50/80/100% thresholds) routed to the team; set TPM limits on every deployment to stop a runaway agent loop or aggressive retry policy from exhausting budget in minutes; tag every deployment with owning team, app, and environment for per-team attribution; and benchmark model quality before scaling to see whether a smaller, cheaper model (e.g., a mini variant) can replace the flagship for that workload. Choosing the right deployment type (Section 3.1) and using Global Batch for offline work are themselves cost levers.
⚠️ Exam Trap: A budget alert does not stop spending — it notifies. If a scenario needs to prevent a runaway agent from burning the budget (not just learn about it later), the real-time control is the TPM ceiling on the deployment, not the budget alert. Picking "set a budget alert" to stop overspend is the trap.
Reflection Question: An agent with an aggressive retry loop could exhaust the monthly budget in an hour. Why is a budget alert insufficient to prevent this, and which control actually caps the burn rate?
