The Standard
The CostMyAI Standard
A four step framework for getting your AI bill under control, where every saving has to be proven, never estimated.
v1.0 · Last updated 3 August 2026
Four levels, in order. What defines each one is the evidence you need to be on it, not the saving it promises. You cannot climb higher than your evidence reaches, and a step you cannot back up with data is not a plan. It is a guess with a price tag on it.
Rung 2 · how it works
The cheapest model that is still good enough.
Not the highest scoring one.
Every benchmark has a margin of error. Take the best model's score, subtract that error margin, and you get the pass mark. Any model at or above it scores so close to the best that the test cannot honestly separate them, so the only thing left to compare is price, and the cheapest one that passes wins. If every model lands inside the error margin, the test proves nothing and we refuse to use it as an argument.
Live example · gpqa tasks (aa:gpqa benchmark)
live prices and scores
Any model scoring 89.02 or higher is close enough that this test cannot tell it apart from the best one. 41 models pass, so price is the only thing left to decide, and the cheapest one wins.
cheapest model that passes
Solar Pro 4
too low · not allowed
as good as the best model, as far as this test can tell
Evidence
What each rung has to prove.
Rung 1 · Compare
Model + price
We know exactly which model you run, and the live per-token price at every host serving those same weights.
Included in every rung above.
Rung 2 · Certify
Benchmark + margin
An independent test score for this kind of task, plus how far that test can be off. No error margin, no claim.
Also proves everything Compare proves.
Rung 3 · Rightsize
Token + complexity profile
Your measured input and output token mix and how hard the tasks actually are, per workload. Not a guess from a price list.
Also proves everything Compare and Certify proves.
Rung 4 · Govern
Full evidence trail
Every automatic action stores the price, the score, the error margin and the condition that would reverse it.
Also proves everything Compare, Certify and Rightsize proves.
The flip side
When the proof is missing, the right answer is no recommendation, never a guess.
No price for what you run today, no score for what you run today, a benchmark that cannot separate the models, nothing cheaper that passes, a speed requirement the alternative would miss, or a saving too small to be worth the switch. In each of those cases we say no and tell you why. That is a feature, not a bug: it is exactly why the recommendations you do get hold up when someone checks them.
Read the full methodology →How to apply this
Two questions show you where you stand.
Answer them honestly. You get the highest step your evidence can back up right now, not the one you would like to be on.
- 01
Can you say what last month's AI bill was spent on, feature by feature and model by model, without logging into a provider dashboard?
- 02
Was your last model change based on a test result, or on a hunch?
Answer both questions and this tells you the highest rung your evidence can actually back up today.
Supporting reading