Actually track what AI is costing you

What does it mean to “track” AI costs?

Tracking AI costs means turning each provider’s raw usage into money at its rates and adding it into one total. Two terms get blurred together along the way. Usage is the raw meter each provider keeps. Spend is that meter priced into money at the provider’s rates. Usage tells you that you sent forty million tokens last week; it can’t tell you what that cost, or let you set it beside a GPU bill counted in entirely different units.

Different kinds of provider meter different things, which is why usage has to be priced into money before any of it can be summed:

Provider type	Example providers	Native unit
Model APIs	OpenAI, Claude	Tokens
Rented GPUs	RunPod	GPU-seconds
Voice synthesis	ElevenLabs	Characters
Hosting / edge	Vercel, Cloudflare	Requests

The number you budget against is spend. So tracking AI costs comes down to three steps, run whenever you want the number: read each provider’s usage, price it into money at that provider’s rates, and add it into one total. The methods below are just different ways of doing those three steps. They vary in how much you do by hand, and how current the number stays.

What are the four ways to track AI costs?

They run from “log in and read it” to “it reads itself.” All four work for somebody. But once you’re a solo dev or a small team running more than two or three providers, the first three stop being worth the effort — and that’s the case this guide is going to make.

1. Provider dashboards

Every provider gives you a usage page — OpenAI’s usage view, Anthropic’s Console, the Gemini billing page — and for a single provider, that’s often enough. You can see what you’ve spent this month and, on most of them, export a CSV.

The limits show up the moment you have more than one. A dashboard is a rear-view mirror: it shows what you’ve already spent, never where the month is heading — and only its own slice, in its own units, with no roll-up. To answer “what am I spending on AI this month” you log into every console and add it up in your head. Fine with one provider. With five it’s a chore, and you’ll make mistakes.

2. A spreadsheet

The natural next step is a spreadsheet: read each console, paste the figure into a row, repeat for every provider, sum the column. It gives you the one thing the dashboards don’t — a total — and you control the layout.

It also has one fatal flaw: a spreadsheet is a snapshot, not a system — which is exactly where a spreadsheet stops keeping up. Keeping it current means re-reading every console and re-typing every figure by hand. Skip a few days and the total goes stale, usually right when a spike is happening. It won’t break spend down by model without more columns, and it gives you no forecast. So you stop updating it, and it drifts.

3. Observability SDKs and AI gateways

The developer-grade answer is to instrument the calls. Observability platforms and AI gateways capture cost per request — some by wrapping your provider SDK, others by routing every request through a proxy between your app and the provider. Done well, you get per-request cost attribution, plus tracing and prompt-level debugging.

That power has a price: these tools live in your code or your request path. There’s an SDK to install, or a gateway that now sits in your latency budget. And they’re built mostly around model tokens, so a GPU box or a hosting bill tends to fall outside what they track. If you want per-request tracing, the trade is worth making. If you just want to know your monthly spend across providers, it’s a lot of plumbing for the question. (For the named tools in this camp and the others — AI cost tools compared by how each collects billing — see the side-by-side.)

4. A billing dashboard

The last option — an AI API cost dashboard — reads each provider’s own usage or billing API directly, when you ask it to, and prices it into money — the same three steps, done in one click. Because it reads the provider’s records instead of intercepting your traffic, there’s no SDK, no gateway, nothing in your request path. It gives you one month-to-date total, a per-provider and per-model breakdown, and a forecast. Each provider documents the feed it reads from — Anthropic’s Admin API, OpenAI’s API, and Google Cloud billing.

This is the approach CostCompass takes, and the rest of this guide shows what it looks like in practice.

Here’s how the four compare:

Method	One cross-provider total	Forecast	Effort to update	Code changes needed
Provider dashboards	No	No	Read each console	None
Spreadsheet	Yes, by hand	No	Re-enter every figure	None
Observability / gateway	Yes	Sometimes	None — passive	Yes — SDK or proxy
Billing dashboard	Yes	Yes	One click	None

How do you forecast where your AI spend is heading?

You forecast it with a run rate: project your recent daily burn across next month. Every method above except the last has the same blind spot — it tells you what you’ve spent, not what you’re on track to spend. Closing that gap is nothing exotic. Take your spend over the last several days, average it per day, and project it across a full month ahead.

That’s how CostCompass forecasts: it takes your trailing seven-day burn rate, multiplies it by the days in next month, and adds next month’s fixed subscriptions. Seven days is long enough to smooth out day-to-day swings but short enough to catch a real change in pace. You get one forward number — at this rate, here’s what next month costs. Because it recomputes from your latest usage every time you refresh, a spike shows up the next time you pull it in, while you can still act on it. Seeing where the spend is heading is the first half; deciding which lever lowers it is the second — see managing LLM costs down for how to act on the breakdown once you have it.

How does CostCompass track your AI costs?

CostCompass is built around the fourth method. You connect each provider once. After that, a click on Refresh reads that provider’s usage and prices it into money, so all those different meters land in one comparable number.

One thing here is intentional: you pull the data, not a background timer. The Refresh button fetches the latest usage from every connected provider on demand. CostCompass doesn’t poll them around the clock — there’s no point calling your providers when you aren’t looking. It stays quiet until you want the number. Then one click brings every provider’s latest usage in at once, already priced and broken down so you can see which provider and model moved it.

Two things make it practical for a single developer. First, nothing touches your code. CostCompass reads each provider’s usage API directly, so there’s no SDK to wrap your calls and no gateway in your request path — your application runs exactly as it did before. Second, your keys are encrypted in your browser before they’re ever stored. Each provider key you connect is sealed with your vault password and saved only as ciphertext CostCompass can’t decrypt — so the App Server only ever holds that ciphertext, and your vault password stays in your browser. When usage is fetched, the key is decrypted in your browser and forwarded to the provider through a broker that holds it for the moment of the call and is built not to log it, so the plaintext stays out of the database and the logs.

It also doesn’t stop at model APIs. The same dashboard rolls Claude and OpenAI up next to the GPU box, the hosting bill, and the voice service — the compute spend most LLM-only trackers ignore. (For how individual providers bill, the per-provider guides go deeper: Claude, OpenAI, Gemini, and the routed models behind OpenRouter, with the full set on the providers page and a wider overview of AI cost tracking across providers.)

Getting started takes three steps:

1. Connect a provider — paste the usage or admin key it gives you. It’s encrypted in your browser before it’s stored, so the server only ever holds ciphertext.
2. Confirm the month-to-date figure matches what you expect for that provider.
3. Add the rest. Each one folds into the same total, forecast, and breakdown — and a click on Refresh pulls the latest across all of them whenever you want it.