Your AI bill jumped and you don't know why. Read the meter in 15 minutes.
Not every AI cost problem is a weekend-long runaway. The more common one is quieter: the bill was $400, now it’s $1,200, and nobody on the team can say what changed. Before you start guessing — or worse, capping the wrong thing — read the meter. Here’s the 15-minute version, in the order that finds the driver fastest.
1. Break it down by model (2 min)
Open your provider’s usage dashboard and group spend by model. This is the single most common cause of a mystery jump: one feature quietly started using a premium model. Maybe a default changed, maybe someone “upgraded” a prompt, maybe a library bumped its default. A frontier model costs 10–30× a small one for the same call, so a single feature switching tiers can move the whole bill. If one model’s line ballooned, you’re probably done already.
2. Break it down by key (2 min)
Now group by API key. If your keys are split by feature or environment (they should be), one key carrying most of the new spend points straight at the culprit — a specific feature, a specific customer, or an abuser using a leaked key. If all your traffic runs through one key, that’s the finding: you can’t tell, and that’s the thing to fix (step 6).
3. Break it down by day (3 min)
Group spend by day and look for the step-change — the date the line jumped from flat to high. Then ask: what did we ship that day? A deploy, a prompt change, a new feature, a marketing push, a new customer onboarded. The bill jumped the day something changed; the date narrows it to a short list.
A tripled bill is almost always one feature, one model, one day.
4. Check the output-token dial (3 min)
Here’s the one people miss. Every call has two dials: how much the model reads (input) and how much it says (output) — and output is billed 3–5× higher. A change that made your model say more — longer answers, a “think step by step” instruction, a bigger response format — can double the bill with no extra traffic at all. If your call volume is flat but spend is up, this is very likely where it went. Check your average output length before and after the jump.
5. Check for a credit that ran out (2 min)
Sometimes the bill didn’t jump — the discount ended. A free trial, a promotional credit, or a committed-spend agreement quietly covered last month, and this is simply the first month you’re seeing the real number. Free credits hide the true cost; their expiry looks exactly like a spike. Check whether a credit lapsed around your step-change date.
6. The real fix: so next month takes 15 seconds, not 15 minutes
If reading the meter was this much work, that’s the actual problem — you’re flying without instruments. The fix is attribution: tag every call by workflow or feature, so your spend dashboard reads “checkout-summary: $600, support-triage: $300” instead of one lump number. Then a jump answers itself: you see which feature moved, the day it happened, on which model. That’s the difference between a monthly investigation and a glance. (how per-workflow attribution works)
If your bill jumped and the five checks above didn’t make it obvious — or you’d rather someone just find it — send it over. It’s just me, so it comes straight to me and I reply fast. Show me your usage and I’ll find what’s driving the increase and set up per-feature attribution so next month you can read it in a glance — free, within a business day.