Google Cloud's spend caps, generally available since October 30, pause new Gemini API, Vertex AI and Cloud Run usage in a project once a monthly budget is hit. They don't stop Compute Engine or storage, they can overshoot, and you pay any overage.
If you build on Google’s AI APIs, you can now put a hard ceiling on what a project spends each month. Google Cloud announced on October 30 that spend caps in Cloud Billing budgets are generally available. A standard budget only sends alerts. A spend cap pauses new usage once the budget is hit.
What changed
Confirmed by Google’s announcement and its spend caps documentation:
- It stops new requests. When a capped project reaches its monthly budget, pay-as-you-go and committed-use requests to that service are halted. Nothing is deleted. Provisioned Throughput (capacity you reserved) keeps serving.
- Four services are covered: the Gemini API, the Gemini Enterprise Agent Platform (formerly Vertex AI), Cloud Run and Cloud Run functions. Compute Engine is not on the list.
- One project, one service per cap. Caps don’t work across folders, organizations, labels or several projects at once.
- It resets monthly. Blocked services come back on the first day of the next month, or you can lift the cap by hand. Lifted services can take up to an hour to resume.
- Alerts go out at 50%, 80% and 100% of the cap to billing admins and project owners.
- AI caps react within minutes, using estimated live costs that now show up in Cost Reports before billing settles.
- There’s an API. You can set, lift or delete caps through the Cloud Billing Budgets API, for example to cap every new sandbox project automatically.
Google’s documentation names no separate charge for the feature.
The limits that matter
Read these before you rely on a cap:
- It can overshoot. Enforcement isn’t instant, and Google says you’re responsible for any overage. It suggests setting the cap a little below your real limit.
- Running resources keep billing. Caps don’t pause fixed usage from persistent resources like compute and storage. Requests already in flight finish and are charged.
- It counts gross cost. Credits and discounts are ignored, so a cap trips earlier than your net bill suggests.
- Some accounts can’t use it. Reseller billing accounts and subscription costs, such as Gemini Enterprise subscriptions, are out of scope.
- You can’t edit a cap while it’s enforced. Lift it first.
What it means
A leaked API key or an agent stuck in a loop can call a model all night. For that case, this is the first real off switch Google has offered: a Gemini API project that stops at $200 instead of an email at 3 a.m. Set a low cap on every sandbox, test and agent playground project now, and a higher one on production with alerts going to someone who reads them. Don’t treat it as total protection. A VM or a storage bucket can still run up charges, and the cap may land a little past your number. If you’re deciding which model to build on, the new Gemini 4 Carbon and OpenAI’s cheaper GPT-6.1 Terra both change the per-request math a cap protects you from.