Priced on the spend we govern, not the savings we grade.
A flat monthly fee, set against the AI spend you put under management. Not a cut of what we save you. Not a charge per head. The model is public; the figure comes in a conversation, and this page explains why both of those are deliberate.
The model, in the open · the figure, in a conversation
How we charge
One flat fee, on the spend under management.
You pay a fixed amount each month, tied to the size of the AI bill you hand us to govern. It does not move when we route a call to a cheaper model, and it does not move when a workflow triples its traffic. You know what Margin costs before you know what it saved. That order is the point: the tool that grades your savings should not have its own paycheck riding on the grade.
How we don’t
Two ways we will never price this.
Both refusals come down to one thing: whether you can trust the number we put in front of you. A pricing model that gives us a reason to inflate it is a pricing model that costs us the only thing we sell.
Never a share of the savings
A grader paid on the number it reports has a reason to report a bigger one.
A vendor paid a percentage of the savings it measures wants those savings to look large. We grade parity ourselves, at a bar we hold ourselves to, on your own tasks, so being paid on our own grade would be marking our own homework. The fee stays flat whether we save you a little or a lot. That is the only version of this that answers the question you should ask any optimiser cleanly: do you make more money if I keep the wasteful agent? No.
Never per seat
We price the spend you govern, not the people you employ.
Per-engineer pricing is what the engineering-intelligence category sells, and it charges you for counting your own team. We are not watching your engineers. We are governing the money your agents spend, so that is what the fee is set against. Hiring a teammate does not change your bill.
Our fee
Flat. The same whether Margin saves you a little or a lot.
A share of savings
The model we refuse. It pays the grader more for reporting a bigger number.
Why it matters
A flat fee has no reason to inflate the savings it reports. That is the point.
A savings-linked fee and an honest grade pull in opposite directions. A flat one does not, so the number we show you is the number we measured.
Why the figure is not here
We have not earned the right to post a number yet.
A company with hundreds of validated customers can publish tiers, because the market has already told it what the product is worth. We have no customers yet, so any figure we posted would be a guess dressed as a fact, and it would anchor every conversation to that guess. A published price is also far harder to move down than a quoted one. So we set the figure against your actual spend, in a conversation, and move it as we learn. Ask, and you get a straight answer, not a runaround.
The model is a promise we can keep in public. The number is a claim we can only make honestly to your face.
See what you would be paying for.
The console runs the whole loop live on real open-source agents: measure, recommend, act, revert, with the parity proof on every change and every number tagged real or simulated. Watch it before we ever talk about a fee.