The invoice is the only number you trust, and it is the last one to arrive. By the time it lands, the call that caused it ran three weeks ago.
Most teams respond by switching to a cheaper model. That moves the bill without telling you which decisions moved with it. The thing to look at first is your own accounting: if a call can happen without being counted, your dashboard is not measuring spend — it is measuring the calls you remembered to instrument.
This page does one thing. It tells you where your LLM spend leaves your books.
If any of those is familiar, the gap is not your model choice. It is that spending happens in places your accounting cannot see, so every decision about cost is made on a number that is already wrong.
The audit does not start with your bill. It starts with your call sites.
Every place your code reaches an LLM is found statically, then each one is checked against four questions: does it record what it spent, does it stop when a budget is reached, does it notice when the model refuses or truncates, and does the task it serves match the model and effort it was given. Call sites that fail the first question are the ones your dashboard has never seen — they are why the invoice and the dashboard disagree.
The output is a map of your call sites, the ones that spend without being counted, and a per-model unit cost table you can put next to the work each call actually does.
Most cost work in this space ends at a report: here is your spend, here are cheaper models, good luck. The report is correct and nothing changes, because the same call site that was invisible last month is still invisible next month.
What is different here is where the work stops. The map is the start, not the deliverable. The top tier ends with a circuit breaker in your codebase and a two-way self-test you can run yourself — a spend ceiling that refuses the call, not a slide that recommends one.
What this is not, stated up front:
Stated plainly: no paying client has run this audit yet. No reviews, no case studies. So the evidence on offer is of one kind only — we ran it on ourselves first, and it found things.
Every number on this page comes from that work. None of it is a client result, and it is not presented as one.
A free mini-audit, one per requester.
Send a public repository URL, or a short description of your stack with roughly how many places call an LLM. You get back the three call sites most likely to be spending outside your accounting, and one of them worked through in full — what it costs, what it fails to record, and the smallest change that makes it visible.
The working limit is 30 minutes. Whatever fits in 30 minutes is what you get, and saying so up front is more accurate for both of us.
The paid version of this, covering every call site with a per-model unit cost table, is listed at $95 (Cost Map). The mini-audit is a sample of that lens, not a discounted version of it.
No credentials. No repository write access. No API keys. Public information and what you choose to describe — that is the whole input.
There is no payment in the mini-audit, so there is nothing to refund. Two commitments instead.
First, if what comes back is something you already knew, reply and say so. Follow-up stops there.
Second, if you move to the paid audit and the direction is wrong, it gets rewritten within the revision count included in that package.
What is not promised: a lower bill. This finds where spending leaves your books and what it costs to bring it back in view. What you do with that — cheaper model, tighter cap, fewer calls — is your decision, and the size of any saving depends on your stack, not on this audit.
The button below opens your mail client. Two lines is enough — a public repo URL or a one-paragraph description of your stack, and the address you want the audit sent to.
Nothing else to prepare. No account to create, no form to fill, no mailing list. The reply goes back to the address you send from.
Request the free mini-auditClosing this page does not cost you the audit. It costs you the interval.
Uninstrumented call sites do not announce themselves. They keep spending at whatever rate they were already spending, and the number you use to reason about cost stays wrong for another billing cycle — which is when the same question comes back, with a larger figure attached.
No fake deadline here. The only real limit is that these are done by hand, so there is a ceiling on how many run at once; past that they are handled in the order received.
The six questions that come up right before people send. Anything not here, add it to the mail.
No. A public repo URL, or a description you write yourself. No credentials, no write access, nothing that would need revoking afterwards.
Describe the stack instead — which provider, roughly how many call sites, and how spend is tracked today. The three most likely leak points are usually identifiable from that.
It tells you where spend is happening outside your accounting. Whether the bill drops depends on what you change after that, and by how much is not something this audit can promise.
That dashboard shows calls it received. It cannot show you the call site in your code that never recorded anything on your side — which is the gap that makes the invoice and your own numbers disagree.
One person, by hand, with the same lens used on our own codebase. No paying client has run it yet — that is stated on this page rather than hidden.
These are handled in the order received and each has a 30-minute working limit. If the queue is long enough to matter, you get told that rather than left waiting.
P.S. — The mini-audit returns three likely leak points and one of them worked through in full, inside a 30-minute limit, with no credentials involved. The real constraint is that a person does each one, so they run in the order they arrive. Send a repo URL or a paragraph about your stack and the address for the reply.