We use essential cookies to make our site work. With your consent, we may also use non-essential cookies to improve user experience and analyze website traffic. By clicking “Accept,” you agree to our website's cookie use as described in our Cookie Policy. You can change your cookie settings at any time by clicking “Preferences.”

LLM Cost Review

Stop paying for the same AI work twice.

We look at what your product asks AI to do, and how it does it. Then we trace the bill to its causes and hand you the fixes in order of what they save.

One run, our own product

  • Work nobody asked for
  • Cache misses
Mike Idengren
Mike IdengrenPrincipal & founder
Keith Mangold
Keith MangoldSenior Advisor

Who does the review

We found this in our own product first. Caching was switched on, and we were still paying full price on every run.

Mike is a Claude Certified Architect, Professional.

Measured on our own product

Same Output. Less Money. Less Waiting.

Cost per run
$0.05Was $0.33. Measured.
Yearly saving
$339,600Modeled at a stated volume.
Read How We Did It

Wait for one build

Measured, one run each.

What a finding looks like

Two Causes. Two Fixes.

CauseCostFix
Product

An Estimate Ran Every Time

Users paid for it and waited for it whether they wanted it or not.

$0.19a run
FixMove it behind a Calculate button. It runs only when someone asks.
Engineering

The Cache Never Matched

Four identifiers nobody asked for made every request look new.

$0.14a run
FixRemove the identifiers. A reused document is now answered from the cache.

The larger saving was a product decision, not an engineering fix.

JarviSIM after the change: the finished process model, with a Calculate button beside Process Metrics that reads Not calculated
The second fix was a button. The model arrives first, and the estimate runs when someone presses Calculate.

You ship the product. We read the bill.

The review

What We Check

What Your Product Asks For

We go through this with the person who owns the product. No tool can decide it for you.

  • Work Nobody Asked For

    AI calls that run on every request when users want the result some of the time.

  • Answers Nobody Reads

    Output that is longer, or made more often, than anyone uses.

  • Model Fit

    Large models doing small jobs, and jobs that need no model at all.

How Efficiently It Runs

We go through this with your developers, in the logs and the code.

  • Cache Misses

    Caching that is switched on but never matches.

  • Repeated Context

    The same long instructions and documents billed in full on every call.

  • Retries and Loops

    Agent loops, retries and timeouts that bill quietly in the background.

Two people going through a printed report at a table

In writing

What You Get

Spend Map

Where the money goes, by feature and by cause.

Ranked Fix List

Each fix with its estimated saving, effort and risk.

Measurement Plan

How to prove each saving, with quality held constant.

Walkthrough

One session with your team to go through the findings.

How it works

Four Steps

  1. Step 1

    Free First Look

    Send last month's invoice. On a short call we tell you whether a review is worth it.

  2. Step 2

    Product Session

    An hour with whoever owns the product: what users need, and what runs anyway.

  3. Step 3

    Logs and Code

    With read-only access, we trace spend to its causes and measure where we can.

  4. Step 4

    Findings

    Written findings and a walkthrough. Fixes are yours to make, or ours on a separate quote.

After the review

A Quarterly Check

Bills drift as features ship. We can come back each quarter, compare your spend with the review and catch new waste early.

This is for you if

  • You run LLM calls in a live product
  • The bill is growing faster than usage
  • Nobody can say which feature costs what
  • Product and engineering disagree about what is worth the cost
  • Users wait on AI responses

It's not if

  • You are still prototyping
  • Your AI spend is too small to matter yet
  • You want a product built from scratch
Reading a printed invoice at a desk beside a laptop

Your part

What We Need From You

  • Last month's LLM invoice
  • An hour with whoever owns the product
  • Read-only access to your request logs or gateway
  • The code that makes the calls

We sign your NDA first.

Questions

Before You Call

Is this only about caching?

No. In our own case the larger saving was a product decision: an estimate that ran for every user moved behind a button. We look at what the product asks the AI to do before we look at how the calls are made.

What access do you need?

Read-only access to your LLM request logs or gateway, and to the code that builds the requests. We sign your NDA first. Tell us what has to stay inside your systems and we work around it.

Do you promise savings?

No. We do not know what is there until we look. If we find nothing worth fixing, we say so.

Will quality drop?

It should not. In our own case the output was the same before and after, and testing required it. Every fix we recommend comes with a way to check that.

Which providers do you cover?

Our deepest work is on Claude. The same checks apply to any provider that bills by token, and to gateways in front of them.

Do you make the fixes?

If you want us to. The review stands on its own, and fixes are quoted separately once you have seen the list.

What happens after the review?

Bills drift as features ship. We can come back each quarter to compare your spend with the review and catch new waste early. Ask about it on the call.

What does it cost?

The first look at your invoice is free. The review is a fixed price, quoted after the call, once we know how many products and providers are in scope.

Bring usyour AI bill.

The first look is free. Have last month's LLM invoice to hand.