Google added pay-as-you-go billing, monthly per-project spend caps, and automatic spend-anomaly detection to Gemini Enterprise this week. The change responds to a problem finance teams are only now discovering: AI budgets built around per-seat licensing don't survive the unpredictable spend of an autonomous agent.
What Changed
Gemini Enterprise now offers pay-as-you-go billing, dropping the fixed monthly fee — companies pay only for actual token and compute usage, at standard API rates, with the per-seat plan still available for those who prefer predictability. Committing to a monthly spend volume earns a 10% token discount on a one-year commitment, or 20% on a three-year one, through what Google calls the Flexible Savings Plan. The most consequential addition for governance is the spend cap: administrators can limit monthly budget per project, get alerts at 50%, 80%, and 100% of that limit, and have agent API calls automatically halt once the cap is hit — unless they choose to manually allow overages. An anomaly-detection feature rounds out the package: it flags projects with abnormal spend and identifies the top three consumption items driving the spike, giving finance and engineering teams visibility into what changed.
Why It Matters
Agent-based AI spend doesn't behave like traditional software licensing. An autonomous agent can multiply API calls within minutes — by design (more tasks, more autonomy) or by mistake (a loop, a poorly scoped task, no limit set) — and a per-seat budget model simply doesn't capture that risk. Google's package is, in practice, applying to agent spend the same discipline infrastructure teams have applied to cloud spend for more than a decade: caps, alerts, cost attribution by project. This connects directly to the McKinsey finding we covered here yesterday — most companies scaling AI agents still can't measure the financial return. Part of the reason is that they also lack clear visibility into where the spend is actually going.
The Impact for Brazil
For Brazilian companies moving from pilot to production with AI agents, the takeaway is practical: treat agent spend like cloud infrastructure spend from day one — with caps, alerts, and per-project attribution — not after a budget surprise. It's worth asking any AI vendor your company is evaluating (not just Google) whether it offers this level of granular spend control; this is quickly moving from a nice-to-have to a basic purchasing requirement for anyone scaling agents past the pilot stage.
Entercast's Take
This launch ties well to what we covered yesterday in McKinsey's survey: the share of companies with real EBIT impact from AI stayed flat even as adoption climbed, and part of the problem is the difficulty of measuring where the value — and the cost — actually sits. Add that to the thread we've followed all month, from DeepSeek's price repricing to AT&T's model routing, and the picture is clear: in 2026, managing AI cost has stopped being about "which model is cheapest" and become its own financial governance discipline, at the level of the agent workload itself. Whoever doesn't build that capability now will discover their own spend the expensive way: on next month's invoice.