Product wants an 80% per-call price cut on your model API — what does that hand a cloner?

## Why this lands on your desk at all An adversary deciding whether to copy your model by querying it runs a two-column comparison: total query spend for the fidelity they need, against acquiring and labelling a comparable corpus plus a training run. **You set and publish one of those two numbers.** That is the only sense in which price per call is a security parameter, and it is a real one: an 80% cut divides the steal column by five and leaves the honest-build column exactly where it was. A competitor's analyst is running this arithmetic off your public price list right now. Your pricing page is the published half of somebody else's ledger. ## What you owe the decision Four things, and they are all numbers or triggers, not objections. **1. The crossing, re-priced.** Re-run the sum at the new price. State the query spend at a *named* fidelity target — the level a reseller in your segment actually needs, which is normally well below matching you — and put it beside the honest-build column with its label line broken out. Say whether the ordering flips. **2. The motive it flips, and the two it does not.** Price moves the resale motive, because that is the only one where both routes end in the same deliverable and money decides. It does not move an adversary who wants a rough stand-in to craft inputs against your production model: that copy is cheap at any price and pays off in effect, not margin. It does not move an adversary using the endpoint as an oracle about the data behind it. Saying "the price cut increases extraction risk" without that split is the sloppy version of this answer. **3. The loss, described honestly.** If a reseller does cross, the outcome is a competitor selling comparable capability at a lower price, trailing your quality and your data refresh. That is margin erosion with a rough time-to-recoup, and it belongs in the same currency as the revenue the price cut is meant to win. It is not a breach, no records leave, and calling it one destroys your credibility in the room where the decision is actually made. **4. A trigger.** The crossing point depends on two prices: yours, and the going rate for the labour that produced your corpus. Say when the estimate gets re-run. ## What you do not get to do Veto. A price is a business instrument and security does not own it. The failure mode here is a qualitative worry — "this makes us easier to steal" — offered with no ledger, which is unanswerable and therefore ignored, and which spends credibility you will want later. The other failure mode is arguing that the price must stay high *as a control*. Holding a price above the market to deter copying is a permanent revenue cost paid to address one motive, and the market will eventually take the decision away from you anyway. If the crossing genuinely matters, the interesting responses are structural rather than a price floor: whether the capability should be sold per call at that price at all, or whether the durable moat is the continuously refreshed corpus and the service around it rather than a snapshot of weights that can be approximated. What specific controls actually bind a determined copier, as against merely raising their bill, is a separate discussion with its own answer. ## When you should not object Be willing to say the cut is fine, and mean it. Two cases: - **The honest-build column is already small.** A commodity task with good public labelled corpora means nobody saves anything by copying you. Price can go to almost nothing without creating a resale motive that was not already there. - **The product is not the weights.** If what customers buy is a service that keeps improving on data that keeps arriving, a snapshot copy trails from the day it is taken and degrades. The clone's ceiling is your behaviour at query time, mistakes included, on the distribution they paid to cover. An analyst who objects to every price cut is not doing threat modelling, they are performing caution. ## What defending the number looks like Expect pushback from both directions, and prepare for both. Engineering will say the model is too good to be approximated from labels; the answer is that the copier does not need to match you, only to clear their own customers' bar on a narrow slice. Finance will say nobody would spend that much on queries; the answer is the ratio between the columns and the payback period, in their units. If you cannot state the sum in currency, with the fidelity assumption written down, you do not have a finding — you have a feeling.