Turn the scare into a system
A surprise bill is not bad luck. It is what happens when cost is variable and nobody is watching it in the right shape. Fix the watching and the surprises stop.
Here are the four habits, in the order you should put them in place. None of them is heavy, and together they close every gap that a bill shock crawls through.
Habit 1: Cap before you launch, not after
A spending cap set after a bad night is damage control. A cap set before launch is insurance, and it is basically free.
Put a per user and per plan limit in place on day one. If you never hit it, it cost you nothing. The one night a bug or a loop tries to run away with your account, it saves you the exact scare you just lived through. There is no version of this where having the cap is worse than not having it. Here is how to set spending caps per user.
Habit 2: Watch cost per user, not just the total
The single total is the number that lied to you. It looks fine right up until the month it does not, because it blends everything together and hides where the money actually goes.
AI cost follows a power law. A few heavy users drive most of your spend, and on a flat plan a couple of them can cost more than they pay. Cost per user puts those outliers in front of you while you can still do something about them, which is the whole idea behind tracking AI costs per user instead of trusting the provider total.
Habit 3: Get alerted in real time, not at month end
The gap between when a cost spike starts and when you find out is where all the damage lives. If you learn about it on the invoice, weeks of overspend already happened.
Close that gap with a real time alert on token velocity, so a runaway user or a loop pings you within minutes. A spike should be a Slack message you get over coffee, not a line item you discover a month later. This is the same move Uber reached for too late in what Uber and Microsoft learned about AI costs.
Habit 4: Review your model choice per feature, quarterly
This is the one habit that does not automate, and it is worth the fifteen minutes. The model landscape moves fast, and prices tend to fall.
A feature you put on a flagship model six months ago may run just as well on something a fraction of the price today. Once a quarter, walk your features and ask whether each one is still on the right model for the job. It is the cheapest recurring win you have, and choosing between a flagship and a mini model is a good place to recalibrate.
Set the first three up once and forget them
Habits are easy to describe and easy to skip. The way they actually stick is to make them infrastructure, so they happen whether or not you remember.
Weckr is that infrastructure for the first three. You wrap your existing OpenAI, Anthropic or Gemini client once, and from then on it tracks cost per user, alerts you on spikes, and enforces the caps you set.
import { Weckr } from '@weckr/sdk'
const wk = new Weckr({
apiKey: process.env.WECKR_API_KEY,
plans: { free: 0, pro: 29 },
})
const response = await wk.chat(openai, {
model: 'gpt-4o-mini',
messages: [{ role: 'user', content: prompt }],
userId: user.id,
feature: 'chat',
plan: user.plan,
})That single wrap covers the cap, the per user cost view, and the real time alert. All you are left with is the quarterly model review, and Weckr even suggests cheaper model swaps to make that easier. Sanity check any feature against the break even calculator while you are at it.
FAQ
How do I avoid surprise AI bills?
Build four habits. Set a spending cap before you launch, not after a scare. Monitor cost per user and per feature, not just the total. Get alerted in real time when usage spikes, instead of finding out at month end. And review your model choice per feature every quarter, since prices and models keep changing. The first three can run automatically, and the fourth is a calendar reminder.
Should I set a spending cap before or after launch?
Before, every time. A cap set after a bad bill is damage control. A cap set before launch is cheap insurance that costs you nothing if you never hit it and saves you a small fortune the one night a bug or a loop tries to run away. There is no downside to having it in place on day one.
Why watch cost per user instead of total cost?
Because the total hides the thing that hurts you. AI cost follows a power law, so a few heavy users usually drive most of the spend, and on a flat plan some of them can cost more than they pay. A single total never shows you that, but cost per user puts the outliers right in front of you before they erode your margin.
How often should I review which model each feature uses?
Quarterly is a good rhythm. New models ship, prices drop, and a feature you put on a flagship model six months ago may run just as well on something a fraction of the price today. A quick review per feature keeps you from overpaying while prices keep falling.
What is the easiest way to set all of this up once?
Wrap your LLM client with a tool that tracks cost per user, alerts on spikes, and enforces caps, so the first three habits run without you thinking about them. Weckr does exactly that in a couple of lines, which leaves you only the quarterly model review to do by hand.
Keep reading
Set it up once so you never search this again
You went looking for this article because something scared you. The goal is that you never need to look for it a second time.
Weckr puts the cap, the per user view, and the live alert in place with a couple of lines, so the next surprise never gets a chance to form. See it running on seeded data at the demo, or read the whole system in the guide to AI cost and margin management.