Claude models on Vertex AI are billed per million tokens against your Google Cloud billing account — pay-as-you-go by default, or a fixed fee if you subscribe to provisioned throughput. There is no built-in spending cap: if a retry loop misfires or a new feature triples prompt sizes, the meter simply keeps running. Google Cloud's answer is the budget-and-alert mechanism in Cloud Billing: define a target amount, scope it to what you want to watch, and get notified as spend approaches it. A budget never blocks requests by itself — it is an early-warning system, and you should treat it as one.
Scope the budget so it means something
An unscoped budget over your whole billing account tells you nothing about Claude. Create a dedicated budget in the Google Cloud console's Billing section (Budgets & alerts) and narrow it two ways: filter to the project(s) where Claude runs, and filter to the Vertex AI service, since Claude charges land under Vertex AI on your bill. If several teams share one Vertex project, service-level filtering can't separate them — that is a labels problem, covered in cost attribution with labels. The cleanest setup, as usual, is one project per team or workload, which turns budget scoping into a project filter. Console flows evolve, so verify the exact clicks against the current Cloud Billing documentation.
Set both actual and forecast thresholds
Budget alert rules fire in two modes, and you want both:
Actual-spend thresholds fire when real accumulated spend crosses a percentage of the budget. A common ladder is 50%, 90%, and 100%. These are your ground truth, but they are inherently backward-looking — by the time 100% fires, the money is spent.
Forecast-based thresholds fire when Google's projection says you are on track to cross a percentage by the end of the budget period. This is the alert that catches a runaway loop on day 6 rather than day 26. A forecast rule at 100% is the single most valuable alert for a token-metered workload.
Route alerts somewhere with an owner
By default, budget emails go to billing administrators — often a finance inbox nobody in engineering watches. Attach the budget to Cloud Monitoring notification channels (email groups, chat integrations, paging) so the team that can actually turn off a runaway job sees the alert. For automation, budgets can also publish to Pub/Sub, letting you trigger your own response — posting to an incident channel, flipping a feature flag, or disabling a noncritical batch pipeline. Be conservative about fully automatic shutoffs: a false-positive spend spike that disables production Claude calls is its own incident.
Size the budget from token math, not vibes
Anchor the number in list prices from Google's pricing page (per 1M tokens, global endpoint): Opus 4.8 at $5 input / $25 output, Sonnet 5 at $3/$15 (promotional $2/$10 through August 31, 2026), Haiku 4.5 at $1/$5, and Fable 5 at $10/$50. Then layer the modifiers that move real bills: regional and multi-region endpoints cost 10% more than global; Google-native batch prediction runs at 50% of on-demand; prompt caching charges cache writes at 1.25x–2x input price but cache hits at only 0.1x; and requests with input context of 200K tokens or more are charged at long-context rates. A budget set from a worked estimate — requests per day × average tokens × price — gives thresholds that mean "something changed," not "we guessed low." The estimation worksheet walks through the arithmetic.
When a forecast alert fires, your first stop is the billing report filtered the same way as the budget, and your second is usage metrics — see the Vertex billing dashboard guide and Vertex billing mechanics. Common culprits, in order: a retry loop without backoff, a prompt that grew past a caching breakpoint, and a traffic shift from Haiku-class to Opus-class models.
Where to go next
Budgets detect; caching and tiering prevent. Read prompt caching on Vertex and model tiering for the two biggest levers on the number the budget is watching.