Mostly metrics is proudly powered by Brex
Every decision a company makes is a financial decision. So finance isn't just the function that reports on the business; it's where the business gets decided.
Which makes it worth asking where your finance team's hours actually go. Reconciling last month. Chasing receipts. Policing a $30 subscription. All of it backward-looking.
That's why I use Brex: corporate cards plus AI-native workflows that absorb the record-keeping. Zero-touch receipt matching, real-time policy enforcement, continuous reconciliation. Controls fire before money leaves the account. Close takes days, not weeks.
Defense handled. Go play offense. Thousands of companies — including Anthropic, Coinbase, and DoorDash — already run on Brex. Head to brex.com/metrics to join them.
How to Analyze Your AI Bill for Savings
👋 Hi, it’s CJ Gustafson and welcome to Mostly Metrics…
Are you ready to save money on tokens?
I’ve got a bucket of these bad larry’s squirrelled away in case of emergency
I’ve had multiple CFOs and FP&A pros come to me and say something along the lines of
“I have no idea what I’m looking at when I get an Anthropic invoice.
I just know it’s going up.”
This is the start of a two part series on managing LLM token costs. All parts will be recorded on video, accompanied by the source files and final outputs.
Part 1: How to Audit Your AI Bill for Savings (this one)
Part 2: How to Budget for Token Costs Internally
In today’s tutorial we start with a real Anthropic invoice + a detailed data export that shows usage by model.
What I've found is there are five common areas for leakage within a company. And you don't have to be technical to find them. In fact, you can use AI to find your own AI waste. Very meta!
Let me show you how I execute this.
First, here’s the full walkthrough.
And here’s everything behind it.
How to try this yourself
Open Claude, drop both files in, and follow along.
The first thing I do is download my invoice.
As you'll see here, it tells me basically nothing. I've had restaurant bills with more detail (I didn’t know they were charging for the diet coke refills).
It has like four lines, which detail the models we used, the totals, and a date. $61,656 for the month. Did we break a window?
The detail exists, it just doesn't come to your email. You need to log into the console and pull it yourself.
Here's what's actually behind that invoice. I know it's an eye sore. But you gotta love meta data and be willing to work with it.
We’ve got Workspace ID, workspace name, API key, the model, and the token counts. Everything the invoice left out that’s actually useful for decision making.
Rather than trying to hack something together with a pivot table, I attach the invoice and the export and hand the whole thing back to Claude.
The first thing I make it do is confirm the two files agree. If the usage export doesn't add up to the invoice total, I'm reading the wrong numbers and everything after this is a waste of time. They match at $61,656. Check.
Then I have it sharpen my question for me before it runs it. That's the prompt on screen.
Fifteen minutes later, sans engineer, and we’ve got some meat.
What I can bet you dollars to donuts is what falls out are solvable blind spots inside your own company.
The first is old models running when new models are cheaper. The tasks weren’t updated when a new model came out.
The second is a lack of caching. You’re getting charged over and over again for words and context that’s already been read.
The third is paying for the fast lane when things can be done at a slower rate in the background.
The fourth is the inverse of the first, which is an expensive model running stuff that a cheaper model can do.
And the fifth is certain jobs consuming most of the bill. Hungry hippos.
And once again, I'm not technical. So I can even use Claude to craft a slack message to send to my IT team to help me change some of the configurations on the backend. I’m the helpful (if not slightly annoying) finance person here to point these things out!
It’s important to note that changing any of them doesn't degrade performance.
I'd encourage you to think of these thematically as you apply them to your own business and look for other blind spots.
The prompts, in order
1. Confirm the files agree, then size up where the money sits.
"You're my finance analyst. Attached are two June files: my Anthropic invoice and my usage export. First, confirm the usage file adds up to the invoice total. Then show me a table: total spend, each workspace as a dollar amount and a percent of the total, and the model and rate type each one uses. Sort by spend, highest first. Don't recommend anything yet."
2. Sharpen the question before you run it.
This is the part most people skip, rushing to ask Claude to get to work. The question you'd type on instinct is usually too vague to get a good answer.
"Before you answer, rewrite my question into a better one. I'm in finance, not engineering. I want specific places I'm overpaying for no reason, each one in dollars per month, each naming the job and the fix, so I can hand it to my engineering team. Give me the improved version first, then run it."
Here's the version it wrote back, which is the one that found the fixes:
"Using the June usage export, find where I'm overpaying with no drop in quality. Check for: a job on an older model when a newer one does the same work for less; a job sending the same large block of text on every request without caching it; a job that isn't time-sensitive running on the standard rate instead of the cheaper batch rate; and an expensive model doing simple, high-volume work a cheaper one could do. For each: name the job, its current monthly cost, the cost after the fix, the monthly savings, and one sentence I can send engineering. Sort by savings, total them, and show the total as a percent of the bill. Use only the numbers in the file."
3. Make it show you the method.
"Now show me the method. Three checks on the same file.
One: my actual average cost per million tokens versus Anthropic's list price, by model, so I can see how much I'm leaving on the table.
Two: what share of my spend is going to each thing (regular input, output, cached text, batch).
Three: how much of the total is sitting in my single biggest job. One short table and one plain sentence for each."
Those three checks work on any usage-based bill. Snowflake, Datadog, and AWS all read similarly.
4. Turn it into a message you can actually send.
"Turn the four findings into one short Slack message to my head of engineering. Start with the total dollars per month and that none of these lower quality. Then list each fix with its job and its monthly savings. Keep it short and friendly. I'm asking, not blaming."
5. Make it a monthly habit.
One bill tells you what's wrong last month. Which is cool, but doesn't help for the future. It also can't tell you if it's getting worse or if anyone fixed it.
"Write me a five-line monthly checklist to run on next month's bill to catch all of this again. Plain language, no jargon. Each line is one thing to check and what makes it a problem."
Download my checklist here. Run it on next month’s bill.
What to change when you run it on your real data
Pull your own export first. My numbers are made up. Yours are in your Anthropic console right now.
The job names are yours to map. Northwind's (my fictional company) are labeled clearly for the demo. Match yours to the real feature or team before you take this to anyone. You’ll look silly if you don’t!
Check before you remove an old model. Ask engineering what still depends on it and why. Sometimes there's a reason. Usually there isn't, lol.
Caching isn't worth it on everything. It has a small cost to set up, so it pays off on large chunks of text you send over and over, not on short one-off requests.
Now that you know how to get data out of your Anthropic console and understand some ways we'll look to trim costs, we're well positioned for next month’s tutorial where we'll go through different ways to budget for AI tokens. Until then!
CJ
