Author
TL;DR: AI price cuts are arriving on printed schedules: promotional clocks on flagship models, expiry dates typed next to holds, increases cancelled before they land. Every one of those cuts flows somewhere. It lands in your margin, on your customer’s invoice, or inside the meter itself, and only one of those destinations is chosen by anyone. Vendors whose unit tracks their model costs donate the deflation automatically. Vendors with a value metric keep the gain and decide where it goes. The difference is which metric you price on, and that choice has to be made before your customers ask.
The last cliff we wrote about ran uphill. The subsidy cliff is what happens when the model layer stops absorbing your inference costs and your unbounded metric turns rising COGS into shrinking margin.
This is the same cliff, facing down. Model prices are falling, in public, on the vendors’ own dated pages. The question nobody has decided is who keeps the difference.
AI Price Cuts Are Scheduled, Not Speculative
The deflation is not a forecast. It is on the vendors’ own pages, with dates.
Google shipped Gemini 3.7 Flash at $0.75 per million input tokens with the expiry printed beside it: the rate holds through December 31, 2026, then roughly doubles. OpenAI put its newest flagship on a promotional clock through late November. Anthropic cancelled a scheduled increase, then removed the dated footnote that had announced it. We log these moves, dated and verified, on the AI Pricing Observatory, and the pattern of the season is concessions with return dates: prices held down visibly, with the reversal already scheduled.
Read those pages the way your customers do. A printed expiry is a repricing event anyone can see coming. A promotional clock tells every buyer downstream of that model that the input cost just moved, and might move back. Deflation with a schedule attached is exactly as destabilizing as inflation, because it hands your customers a number to negotiate with and a date to do it on.
To be precise about what this is not: this is not a bubble argument. Whether AI capital markets are overheated is somebody else’s debate. Model-layer price cuts are documented commercial events, and they reprice your cost base whether or not anyone’s valuation thesis survives.
We have watched this conversation before. When inflation cooled, the buyers who had absorbed cost-justified increases came back at the next renewal and asked, reasonably, whether the increase would now be reversed. A price change justified by your costs is an invitation to renegotiate on your costs, and the invitation runs in both directions. Model-price deflation hands that same question to every customer of every vendor that has ever mentioned inference costs in a pricing conversation.
Three Places the AI Price Cut Can Land
When your model bill drops, the savings flow to one of three destinations. Two of them happen by default. Only one is a decision.
Capture 1: Your margin
If your pricing carries a value metric, anchored to what customers receive rather than what models cost, the cut lands in your gross margin. It stays there until you decide otherwise. This is the position every vendor thinks it holds. Fewer actually do.
Capture 2: Your customer’s invoice
If your unit is cost-linked, tokens passed through, compute metered at cost-plus, credits pegged to provider rates, the deflation flows to customers automatically. You donated it. No one in your company decided to cut prices, and yet prices fell. The difference between a pricing model and a value metric is exactly this: the metric decides what your revenue tracks. A metric that tracks your costs makes your customers the beneficiaries of your suppliers’ price war.
Capture 3: The meter itself
Between those two sits the quiet option. A vendor that prices in credits can change the conversion table, the multiplier between the unit customers buy and the work it purchases, and reprice everything while the published price list holds perfectly still. In a credit system, the real pricing lever is the conversion table, how much work each credit buys, rather than the published rate; the lineage of minted currencies explains why. Your multiplier policy needs a defensible answer at renewal, decided before anyone asks for it.
Cost-Linked Units Donate the Deflation
The asymmetry runs one way. When model prices rose, cost-linked vendors had a painful but defensible conversation: our inputs became more expensive. When model prices fall, those same vendors have no conversation at all. The unit does the repricing for them, in the customer’s favor, automatically.
Automatic pass-through is not a policy; it is the absence of one. A unit that tracks vendor cost hands the customer every improvement your suppliers deliver, which means your R&D negotiations, your committed-spend discounts, your migration to a cheaper model, all of it flows through. You did the work of capturing the savings and then the meter gave them away.
The billing-layer argument frames automatic pass-through as fair by design. That is the wrong frame. Fairness is a deliberate allocation with a reason attached; a default nobody chose hands the customer the benefit of your supplier negotiations for free. And a cost-linked unit that moves visibly with provider announcements signals your repricing calendar to every buyer downstream. That visibility invites timing games the vendor cannot control; a value metric removes the signal entirely.
A value metric holds the gain in one place, on purpose, where it can fund a deliberate answer: cut, reinvest, or bank the margin. Each is legitimate exactly once it is chosen rather than defaulted into.
Are Your Units Wired to Pass Deflation Straight to Buyers?
If your value metric tracks compute or token consumption, model-layer price cuts flow through to your customers automatically. Find out whether your licensing, packaging, and pricing architecture is structurally exposed to that asymmetry.
The Renewal Question Is Already Scheduled
Your customers’ procurement teams read the same announcements you do. The question arrives at renewal, in some form of: your model costs fell this year, why didn’t our price?
The vendors who struggle with that question are not the ones who kept the margin. They are the ones who never decided anything and have to reconstruct an answer under deadline. Vendors who answer it well already know which of three positions they hold: a cut with a schedule, a reinvestment with visible improvements, or a held price with the value it reflects. Any of the three survives contact with procurement. What does not survive is discovering your own answer during the negotiation.
This is the architecture question underneath every AI pricing decision: the unit determines which conversations are even possible. Decide the unit and the deflation question has an owner. Inherit the unit and the deflation answers itself, usually against you.
In renewals where the vendor holds a value metric that captures what the software returns to the customer’s business, the cost question never reaches the table. That position is not available to every product at every stage. When it is, your costs are none of the buyer’s business, and the R&D you invested to lower them stays in your margin. The path there runs through licensing, packaging, and pricing, in that order.
Tests to Run Before the Question Arrives
Four diagnostic questions determine whether your pricing architecture captures AI cost deflation or donates it automatically.
- If your largest model input price fell 30% tomorrow, which line moves: your margin, your customer’s invoice, or your meter’s conversion table? If you cannot answer in one sentence, the meter will answer for you.
- Can your effective price change without a customer-visible event? If yes, you hold the quiet lever, and so does every competitor who prices the same way. Decide your policy for it before someone asks whether you have one.
- Does your paper say anything about cost-linked adjustment, in either direction? An agreement that is silent on falling input costs was priced for a world where they only rose.
- Who in your company owns the decision of where deflation lands? A name, not a committee. If the answer is nobody, the decision defaults to your value metric, the meter itself, and the meter does not work for you.
If you are staring at a rate card that moves on someone else’s schedule, talk to a pricing expert and describe your unit before your customers describe it for you. A pricing expert replies, not a form.
