How to Set Spending Limits and Alerts on Pay-As-You-Go AI Tools

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Set Spending Limits and Alerts on Pay-As-You-Go AI Tools.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Set Spending Limits and Alerts on Pay-As-You-Go AI Tools.

Set a hard monthly limit wherever the tool offers one and alerts everywhere else: spend limits on OpenAI and Anthropic API projects, a capped auto-recharge on prepaid credits, monthly caps on extra usage in ChatGPT and Claude, and a deliberate choice about Zapier's pay-per-task and Make's credit auto-purchase. Then check usage weekly.

Alerts and limits are different things, and many tools default to alerts or to paying the overage. An alert emails you and keeps spending; a hard limit stops the service, which can break a customer-facing workflow at 9pm on a Saturday. So for each tool, decide which is worse: an unexpected bill or a stopped automation. For an internal weekly report, stopping is fine. For booking confirmations, you'd rather pay a little more and be told.

Follow me on Instagram@sagnikteaches

Alerts, caps and hard stops do different things

ControlWhat happens when it's reachedGood for
AlertYou get an email or notification; spending carries onAnything customer-facing, as an early warning
Capped top-upPrepaid credits reload automatically, but only up to a monthly total you setAPI accounts that must keep running, within a known ceiling
Overage billingUsage continues at a higher per-unit price, sometimes up to a ceilingBusy months you'd rather pay for than miss
Hard stopRequests fail or workflows pause until the month resets or you raise the limitExperiments, internal tools, and a last line of defence

The strongest setup layers them: an alert at about half the limit, a second at 80%, and a hard stop well above a normal month, so the stop only ever meets a genuine runaway.

Connect on LinkedInSagnik Bhattacharya

Decide per tool: stop the service or pay the overage

List every tool that bills by usage, then put each one in a column. A quick way to decide: ask what a customer experiences if it stops for 12 hours.

Subscribe on YouTube@codingliquids
  • Stop is acceptable: the monthly report generator, a content-drafting script, a test chatbot, an AI step that summarises internal notes. Set hard limits.
  • Stop is costly: the website chatbot, the automation that confirms bookings, anything that replies to customers. Set alerts early, allow overage to a ceiling, and give the tool a fallback for when it does stop.
  • Stop is invisible: automations where a failure produces no error anyone sees. These are the dangerous ones; add a daily check as well as a limit. Stopping Zapier and Make automations breaking silently covers the monitoring side.

Where each vendor keeps its limits, as of September 2026

Settings move around, so treat the labels as a guide to what to look for.

ToolWhere to find itWhat you can setWorth knowing
OpenAI APIThe organisation's limits page, and each project's Limits tabA monthly spend limit, with an "Enforce a hard limit" toggle; spend alertsHard limits return a 429 error; enforcement isn't instant, so spend can slightly exceed the limit
OpenAI prepaid creditsBilling settingsAuto recharge threshold and amount (minimum $5), plus an optional monthly recharge limitPurchased credits expire after a year; usage halts at $0 if a recharge fails
Anthropic API (Claude Console)Settings, then Billing, then Spend limits; each workspace's Spend limits tabAn organisation limit below your tier's cap; workspace limits with alert thresholdsTier caps are $500 (Start), $1,000 (Build) and $200,000 (Scale) a month; no limits on the Default Workspace
Anthropic prepaid creditsSettings, then BillingAuto-reload: a minimum balance and the amount to reload toCredits expire a year after purchase; at zero, API calls stop
ChatGPT extra usageSettings, then Usage (personal plans); Workspace settings, then Billing (Business)Auto top-up with a minimum and target balance; on Business, a Monthly recharge limit and per-user limitsLeaving Business's Monthly recharge limit blank allows unlimited automatic reloads
Claude extra usage (Pro and Max)Settings, then UsageA monthly spending cap, auto-reload, or "Set to unlimited"Billed at standard API rates, separately from the subscription
ZapierBilling and usage, then Billing settings, Per-task billingPay-per-task on or off (account owner only on Professional and Team)On by default for accounts created after January 2024; runs to 3x your plan's tasks at 1.25x (annual) or 2.5x (monthly) the base rate
MakeOrg in the sidebar, then the Subscription tab"Enable credits auto-purchasing"; per-team credit limits on EnterpriseExtra credits cost 25% more and auto-purchase buys 10,000 at a time; running out raises a fatal error that switches off scenario scheduling
n8n cloudPlan settingsChoose a plan with headroom; monthly Pro-2 subscribers can add executionsAt the limit, executions fail immediately and nothing queues; limits reset on the 1st of each month
Gemini API via Google CloudCloud Billing, BudgetsAlerts-only budgets; spend-cap budgets (preview) where the service supports themAn alerts-only budget never stops spending on its own

OpenAI's spend limits guide and Anthropic's rate and spend limits page are the two pages to bookmark; both describe exactly what an application sees when a limit bites. One Make detail catches people out: because running out of credits switches scenario scheduling off, topping up doesn't restart anything by itself. After you buy credits, check each scenario is switched back on.

Setting it all up in about an hour

  1. Find every usage-billed tool (10 minutes). Search the last three months of card and bank statements for AI and automation charges. This finds the forgotten ones: a test API key a freelancer set up, a trial that rolled into a paid plan.
  2. Mark each one stop or pay (5 minutes), using the 12-hour question above.
  3. Split API use into separate projects or workspaces (15 minutes). Give each automation its own OpenAI project or Claude Console workspace, named for what it does, such as "Website chatbot" or "Monthly report". A limit then stops only the workflow that ran away, not every AI feature you have. Anthropic doesn't allow limits on the Default Workspace, which is another reason to create named ones.
  4. Set limits and alerts on each project or workspace (10 minutes), then cap any prepaid auto-recharge.
  5. Cap extra usage in the chat apps (5 minutes), or leave auto top-up off entirely.
  6. Choose the overage behaviour on Zapier, Make and n8n (10 minutes).
  7. Route and test the alerts (5 minutes), as described further down.

A barber shop's setup: expected spend, caps and alerts

An illustrative two-chair barber shop uses three pay-as-you-go tools. A Zap on Zapier Professional (750 tasks, $29.99 a month on monthly billing) drafts replies to booking enquiries and Google reviews, calling OpenAI's API with the shop's own key. The owner has ChatGPT Plus at $20 a month. And the API runs on prepaid credits.

First, the expected spend. The Zap runs about 180 times a month at 3 tasks a run, so 540 tasks, comfortably inside 750. Each AI call uses about 1,200 input and 250 output tokens on gpt-6-luna ($0.10 and $0.50 per million tokens), which is about $0.00025 a call, or roughly $0.05 a month. The API bill is tiny; the risk isn't the normal month, it's the abnormal one.

Here's what the owner set, in about 40 minutes:

ToolExpected monthControl setWhy that number
OpenAI API project "shop-automations"About $0.05Monthly spend limit $10 with hard limit enforced; alert at $3Far above normal, far below painful; a runaway stops at $10
OpenAI prepaid creditsn/a$20 of credits; auto recharge $10 when below $5, monthly recharge limit $20The account can never buy more than $20 of credit in a month
Zapier540 of 750 tasksPay-per-task left on; 80% and 100% emails forwarded to the owner's phoneBooking replies matter more than a small overage
ChatGPT Plus$20No extra credits bought; auto top-up left offThe subscription is the whole budget

The API limit looks absurdly generous at 200 times a normal month. It's deliberate: the point is to cap a disaster at $10, not to police pennies, and a limit set too tight trips in a busy week and stops the booking replies.

The Zap that looped overnight, and which control caught it

Three weeks later, the owner added a step to the Zap that logged each enquiry to a Google Sheet. Another Zap, set up months earlier to tidy that sheet, triggered on new rows and added a formatted copy to the same sheet. The two now fed each other.

A loop doesn't error, so Zapier's safeguards for failing Zaps never fired; every run "succeeded". Here's how the night went, illustratively:

  • 10:40pm: the loop starts, using about 4 tasks a cycle across the two Zaps.
  • 11:05pm: the account passes 80% of its 750 tasks; Zapier's email arrives, but the owner is asleep.
  • 11:20pm: the plan's 750 tasks are gone and pay-per-task takes over, because it was left on.
  • Next morning: usage stops at the pay-per-task ceiling of three times the plan, 2,250 tasks.

The quick sum: 1,500 extra tasks at 2.5 times the base rate. On the monthly 750 plan, the base rate is about $0.04 a task ($29.99 divided by 750), so each extra task costs about $0.10, and the overnight loop added roughly $150. On annual billing, at 1.25 times a lower base rate, the same loop would have cost about $50. And because the Zap's AI step used the shop's own key, the OpenAI project's $10 hard limit capped that side of the damage, as intended.

The fixes, in order of importance: a free filter step so the tidy-up Zap ignores rows written by the enquiry Zap; pay-per-task switched off, so future runaways are held for replay instead of billed; and a note in the Zap descriptions saying which sheet each one writes to. When Zapier gets too expensive covers the wider question of whether a busy account should move tiers, or platforms.

A chatbot retry loop on the Claude API, priced

API runaways look different from Zap loops, and they can cost more per hour. Take an illustrative personal trainer whose website chatbot was built by a freelancer on Claude Sonnet 5 ($2 per million input tokens, $10 per million output). A bug made the widget retry after every slow reply, and each retry resent the whole conversation plus a 15,000-token FAQ.

Each call carried about 16,000 input tokens and produced about 300 output tokens: roughly $0.032 plus $0.003, so $0.035 a call. One stuck browser tab retrying every few seconds overnight can make 3,000 calls, which is about $105 by morning, for a chatbot that normally costs a few dollars a month.

The trainer had put the chatbot in its own Claude Console workspace with a $25 monthly spend limit and an alert at $10. The alert email arrived at 1am; the limit stopped the calls at about $25, and the widget showed its fallback line ("Sorry, I can't answer right now. Leave your email and we'll reply this morning."). When a limit you set is reached, Anthropic's API returns an error saying the specified workspace usage limit has been reached, which is exactly what the fallback was built to catch. The freelancer fixed the retry logic the next day, and the limit never needed raising.

Picking the cap figure without guessing

Use three months of real usage if you have it; if not, run for two weeks with alerts only and multiply up. Then apply a simple rule:

  • Customer-facing tools: hard limit at about 2 to 3 times a normal month; alerts at 50% and 80% of the limit.
  • Internal tools and experiments: hard limit at about 1.2 times a normal month; one alert at 80%.
  • Tiny API bills: set a fixed ceiling that would hurt only slightly, such as $10 or $25, rather than a multiple of pennies.
  • Prepaid credit reloads: always set the monthly recharge limit where the tool offers one; an uncapped auto-reload is a credit card with no limit.

Revisit the numbers every quarter, and whenever you add a workflow. If you're still working out what the whole AI budget should be, how much a small business should budget for AI gives ranges by business size, and what the ChatGPT API costs for a business automation shows how to estimate per-call costs before you build.

Send alerts somewhere a person will see them

An alert that lands in the inbox of whoever set up the account three years ago is not an alert. Three rules:

  1. Use a shared address for billing and usage emails on every AI tool, such as accounts@ your domain, and forward it to the owner's phone.
  2. Name one person who acts on alerts, and a second for holidays.
  3. Test the route. Temporarily set an alert threshold just above today's usage, let it fire, and confirm the right phone buzzes. Then set it back.

Also check who holds the owner role. On Zapier Professional and Team, only the account owner can switch pay-per-task on or off. If that person has left, you can't change the setting when you most need to.

A five-minute Monday check

Limits catch disasters; a weekly glance catches drift. Every Monday, open each usage page and compare with last week:

  • OpenAI and Claude Console usage: any day more than three times the average?
  • Zapier task usage: on track to finish the month under the plan?
  • Make credit usage: any scenario using far more than usual, and all scenarios still switched on?
  • n8n executions: headroom left for the rest of the month?
  • ChatGPT and Claude extra-usage balances: any top-ups you didn't expect?

Write the five numbers in a sheet. After a month you'll know what normal looks like, which is the only way to spot abnormal early. If a single automation dominates the numbers, that's a prompt to look at its design, and Zapier vs Make vs n8n for AI automation can tell you whether it belongs on a platform that counts its steps differently.

Spending limits on AI tools: common questions

Does a ChatGPT Plus or Claude Pro subscription cover API use?

No. ChatGPT Plus and Pro and Claude Pro and Max don't include API calls. The API is billed separately, per token, on each company's developer platform, so an automation using your own API key runs up its own bill. That's why the API account needs its own spend limit even if your chat subscription is a fixed monthly price. The extra-usage credits inside the chat apps are a separate meter again.

Will a hard spend limit break my customer chatbot?

It can. When OpenAI enforces a hard limit, affected requests fail with a 429 error; when an Anthropic organisation hits a limit you set, requests fail with an error until the limit is raised or the month resets. Any customer-facing tool behind that key stops answering. Give the tool a polite fallback message, set the limit at around twice a normal month, and put an alert at 50% so you hear about a spike well before the stop.

Why was I charged more than the limit I set?

Usually one of three reasons. OpenAI says enforcement isn't instantaneous, so recorded spend can slightly exceed the limit while it takes effect. Zapier's pay-per-task keeps running up to three times your plan's tasks, and that is by design, not a breach. And Make's credit auto-purchase buys 10,000 credits at a time, so one purchase can go well past what you needed. Check each tool's overage setting, not just the headline limit.

Further reads

Sources: OpenAI API docs on spend limits and OpenAI help articles on prepaid billing and ChatGPT credits; Claude Platform docs on rate limits, spend caps and workspaces, and Claude help articles on API payment and usage credits; Zapier help on pay-per-task billing and task usage; Make help on extra credits, credits per team and errors; n8n support on execution limits; Google Cloud Billing docs on budgets. Checked September 2026.

Want caps and alerts set across all your AI tools?

On a 1:1 call we'll list every pay-as-you-go AI tool you use, decide where a hard stop is safe and where an alert is better, and set the numbers from your real usage.

Book a 1:1 call with me