Key Takeaways
- A budget alert sends email while the meter keeps running; only a hard cap stops new usage.
- From Hong Kong, use Gemini through Google Cloud, where a spend cap budget can pause the service per project.
- Zapier and Make keep running past your plan unless you switch off pay-per-task or credit auto-purchasing.
- No cap is instant: Google and OpenAI both say overage can still accrue for minutes.
- Set each cap about 10% under your real limit and record it in one shared register.
In February 2026 a three-person team that normally spent $180 a month on Google Cloud woke up to $82,314.44 in Gemini charges run up in 48 hours with a stolen API key, according to The Register.
Most teams have a budget alert. Few have a cap that would have stopped that bill. Here is the fix, tool by tool.
Why is a budget alert not enough to stop a runaway AI bill?
A budget alert only sends a notification; API traffic keeps flowing and you keep paying. A hard spend cap blocks new requests once the amount is reached. OpenAI's own documentation draws this line explicitly, and Google Cloud separates "Alerts only" budgets from "Spend cap enforcement" budgets for the same reason.
OpenAI's spend limits guide puts it plainly: spend alerts "do not enforce a cap". Google Cloud's spend cap documentation treats alerts-only budgets as a separate product from budgets that pause a service. So the first job is to find out which kind of control each of your tools actually has.
What do you need before you set AI spend caps?
You need owner or billing-admin access to each tool, a list of every place your team pays for AI by usage, and one monthly number per tool. Fixed per-seat subscriptions cannot run away, so focus on API keys, automation platforms and anything with auto top-up or overage billing switched on.
You need
- Admin rights: Billing Account Administrator or Project Owner on Google Cloud, account owner on Zapier, organisation admin on Make and Perplexity
- A usage-billed tool list: API keys, Zapier, Make, n8n, Perplexity API, any AI Studio projects used by overseas colleagues
- One monthly ceiling per tool, in US dollars, agreed with whoever signs the invoice
The comparison table to keep (last verified 9 November 2026)
| Tool | Real hard cap? | Where to set it | Usable in Hong Kong? |
|---|---|---|---|
| Gemini on Google Cloud (Gemini Enterprise Agent Platform, formerly Vertex AI) | Yes: spend cap budget pauses the service (preview) | Billing > Budgets & alerts > Create new budget > Spend cap enforcement | Yes, including the Hong Kong region |
| Google Cloud "Alerts only" budget | No, email alerts only | Billing > Budgets & alerts > Create new budget > Alerts only | Yes |
| Gemini API in Google AI Studio | Yes, but experimental with about 10 minutes of lag | AI Studio > Spend > Monthly spend cap > Edit spend cap | No, Hong Kong is not on the supported list |
| Perplexity API | Only if auto top-up is off (requests fail at a zero balance) | API Console > Billing > Auto top-up | Yes |
| n8n Cloud Starter and Pro | Overage fees listed as "Capped" | Usage shown in Insights; plan on n8n.io/pricing | Yes (hosted in Frankfurt) |
| Zapier | Yes, if pay-per-task is off; otherwise stops at 3x plan tasks | Settings > Billing and usage > Billing settings > Per-task billing | Yes |
| Make | Yes, if credit auto-purchasing is off | Org > My Plan > Subscription > Enable credits auto-purchasing | Yes |
How do you set a hard spend cap for Gemini from Hong Kong?
Use Gemini through Google Cloud, not AI Studio, because Hong Kong is not on AI Studio's supported list. In the Cloud console, create a new budget, choose Spend cap enforcement, scope it to one project and the Gemini Enterprise Agent Platform service, then enter a target amount. New usage pauses when the estimate passes it.
Google's available regions page (updated 22 October 2026) lists Taiwan, Japan and Singapore but not Hong Kong, so the AI Studio spend cap is only relevant to colleagues working from elsewhere.
Step 1: Create the spend cap budget
- Google Cloud console > Billing > Budgets & alerts > Create new budget
- Define: select Spend cap enforcement (not Alerts only), name it, for example "gemini-prod-cap"
- Scope: time range is fixed to Monthly; pick one project and one service, Gemini Enterprise Agent Platform (aiplatform.googleapis.com)
- Amount: enter the target amount, set a little below your true ceiling
- Actions: alerts at 50%, 80% and 100% go to billing admins and project owners automatically; click Finish
Step 2: Check the status column
On the Budgets dashboard the Spend cap status should read Configured. It changes to Enforced when the cap trips, and back to Configured on the first of the next month. If you already have an alerts-only budget for that project, you cannot convert it; delete it and create a new spend cap budget.
How do you stop Zapier, Make and n8n from billing past your plan?
Switch off the overflow billing. In Zapier, turn off Per-task billing so Zaps are held at your plan limit. In Make, disable credits auto-purchasing so scenarios stop when credits run out. On n8n Cloud Starter and Pro, overage is listed as capped, so watch execution counts in Insights.
Step 3: Zapier
Settings > Billing and usage > Billing settings tab > Per-task billing toggle off. Only the account owner sees this on Professional and Team plans. Zapier's help page says pay-per-task is on by default for accounts opened after January 2024 and, when on, can take you to three times your plan's tasks: a Professional 750 plan can run 2,250 tasks in a month.
Step 4: Make
Org (left sidebar) > My Plan > Subscription tab > turn off Enable credits auto-purchasing. Make's help centre explains that auto-purchasing buys 10,000 credits at a time, at a 25% premium, up to the size of your plan again.
Step 5: n8n
n8n bills per workflow execution, not per step. The pricing page lists overage fees as "Capped" on Starter and Pro. The cost that runs away inside n8n is usually the model API key you plugged into an AI Agent node, so cap that key at its provider as well.
How do you cap Perplexity API spending?
Perplexity's API runs on prepaid credits, so the balance itself is your cap if auto top-up is off. Requests return a 402 error at zero. With auto top-up on, the documentation describes no monthly ceiling, so buy a fixed amount instead and top up by hand when the dashboard tells you.
Step 6: Perplexity
API Console > Billing > Auto top-up: switch it off, then buy a fixed credit amount. The Perplexity API FAQ confirms that an empty balance returns 402 Payment Required, and that admins can filter usage by API key on the same Billing page. One key per workflow makes it obvious which automation is spending.
What is the worst-case monthly cost once caps are on?
Your cap plus a short overage window. Google says AI Studio caps lag about 10 minutes and Cloud spend caps are not instant; OpenAI says recorded spend can slightly exceed a hard limit. The caps themselves cost nothing to set, so the only trade-off is that a capped workflow stops mid-month.
Work out each tool's worst case with and without the overflow switch:
| Example plan | Overflow on | Overflow off |
|---|---|---|
| Make, US$9 plan with 10,000 credits | Up to 10,000 extra credits at US$0.001125 each: up to US$20.25 a month | US$9, scenarios stop at 10,000 credits |
| Zapier Professional, 750 tasks | Up to 2,250 tasks billed | 750 tasks, Zaps held |
| Gemini on Google Cloud, US$100 ceiling | Alerts only: no limit at all | Cap at US$90 to absorb reporting lag |
Figures from Make's and Zapier's own help pages; the 10% Google buffer follows Google's advice to cap "slightly below your absolute limit".
What goes wrong with AI spend caps, and how do you fix it?
Three things: overage during reporting lag, caps that measure list price rather than your discounted bill, and long-running jobs that finish after the cap trips. Fix them with a buffer below your real limit, restricted API keys, and a weekly check of each tool's usage page instead of relying on the cap alone.
Error 1: "I set a cap and still got charged more"
Google's Gemini billing documentation says AI Studio project caps are experimental and you remain liable for overage during roughly 10 minutes of billing lag; batch jobs and agent sessions can run past the cap. Fix: set the cap 10% low and avoid batch jobs on capped projects.
Error 2: "The cap tripped earlier than my invoice suggests"
Google Cloud spend caps use gross, estimated cost at list price, without credits or discounts. Fix: if you hold credits, set the cap against list price, not what you expect to pay.
Error 4: "A leaked key ran up the bill"
In the US$82,314 case, the team's usual spend was US$180 a month. A spend cap set at US$200 would have limited exposure to roughly US$200 plus a few minutes of lag. Fix: restrict every key to the APIs it needs and rotate any key that has been in client-side code.
Keep a cap register
AI SPEND CAP REGISTER (review every Monday)
Tool: [tool name]
Account owner: [name]
Cap type: [hard cap / alert only / prepaid balance]
Monthly cap (USD): [amount]
True ceiling (USD): [amount]
Where set: [exact menu path]
Overflow billing: [off / on, reason]
Last verified: [date]Your 30-minute spend cap checklist
- List every usage-billed AI tool and its account owner
- Gemini via Google Cloud: one Spend cap enforcement budget per project
- Zapier: Per-task billing off, or accept the 3x ceiling knowingly
- Make: credits auto-purchasing off
- Perplexity API: auto top-up off, fixed prepaid balance
- Every cap 10% below the true ceiling, recorded in the register
Alerts tell you about a problem. Caps put a ceiling on it. Last verified: 9 November 2026; menu names change, so recheck the links above before each quarter's budget review.
Empowering AI. Centered on you.
Rather not babysit caps across five tools?
If you don't want to build and maintain these workflows and guardrails yourself, hand them to AI employees that UD deploys and runs by role, with $0 setup and a free AI employee configuration assessment. We'll walk you through every step, from picking the first role to day-to-day maintenance.
Get a Free AI Employee Assessment
Frequently Asked Questions
Can I set a hard spend limit on the OpenAI or Anthropic API?
Yes. OpenAI offers Enforce a hard limit under Project settings > Limits, and Anthropic's Console sets monthly organisation and workspace limits. Neither provider officially supports Hong Kong, so check your company's policy first.
Does a Google Cloud spend cap stop requests already running?
No. Google's documentation says in-flight requests finish and are charged. Only new usage of the capped service in that project is paused.
Will a spend cap delete my data or workflows?
No. Caps pause usage; Zapier holds Zap runs you can replay later, and Google Cloud resumes the service on the first of the next month or when you lift the cap.