Skip to content
Aero Interactive

Your AI feature's margin moves every time the model's price does

By

Buyers want consumption pricing for AI, but not for the rest of the product

A tracking survey of 830 global IT decision-makers found buyer preference for consumption-based pricing on generative AI features rose 5.3 percentage points, to 42.9 percent, from 2H 2025 to 1H 2026, according to Futurum Group.

The same tracking survey found preference for consumption pricing on core software platforms fell 5.8 percentage points, to 30.1 percent, over that same wave-over-wave comparison. Buyers want to pay for AI by what they use. They want the rest of the product to stay predictable.

That split looks like a pricing decision. It is also a margin decision. Consumption pricing only protects margin if the vendor knows what the consumption actually costs.

A consumption price set lower than the actual cost per unit is not a pricing model. It is a subsidy. It compounds with every heavy user the feature attracts.

A model upgrade can move your cost per query by hundreds of times, without anyone touching the price

As of November 2026, OpenAI's own published API pricing lists per-million-token rates for its standard-tier text models. Input tokens range from $0.05 to $30.00 per million tokens, a 600-fold spread, depending on which model answers a query. Output tokens range from $0.40 to $180.00 per million tokens across that same lineup, a 450-fold spread.

A feature can move between those rates with no change to its own code. Picture a team that ships an AI summarization feature inside the existing subscription price. At launch it runs on a low-cost model and costs a fraction of a cent per summary.

Six months later the team upgrades the model for better output quality. The same feature now costs several cents per summary at the heaviest decile of usage. Nobody changed the price. The margin moved on its own.

A provider's price list is a cost input the product team rarely monitors. A model upgrade is a pricing event even when no price on the product side has changed.

That gap matters most for teams that priced the feature once, at launch, and have not revisited it since. A provider's per-token price list changes on its own schedule. It rarely matches the product team's release calendar.

A decision rule: price for the usage tail

Average cost per user looks fine until a small group of heavy users runs the feature constantly. What matters is the cost at the 95th percentile of usage. Calculate it against the most expensive model the feature could reasonably run on, including a model it might be upgraded to later.

If that number sits inside the margin budget, the feature can stay inside the base price. If it does not, the feature gets metered, capped, or priced separately before launch. Waiting until a provider's price change forces the question costs more than running the test up front.

A cost ceiling set this way also gives the team a clear answer when a provider announces a price increase, instead of a scramble to recalculate margin after the fact.

Production readiness for an AI feature is not only about accuracy and reliability. Our note on moving an AI pilot into production covers the rest of that checklist.

Before you price an AI feature, check:

  • Has this feature been priced at the 95th percentile of usage, not the average user?
  • Has it been priced against the most expensive model it could reasonably run on, not only the model it runs on today?
  • Is there a defined trigger for what happens when a provider changes its per-token price?
  • Does someone own a monthly review of this feature's margin, the way a usage-based company reviews cloud spend?

FAQ

Does usage-based pricing solve the margin problem on its own

No. Usage-based pricing moves cost risk toward the buyer. That only works if the price per unit is set correctly. A usage-based price set before anyone modeled the 95th-percentile cost carries the same risk as a flat price. It just adds a billing step.

What if the feature routes to different models automatically

Price against the most expensive model the routing logic could select. Do not price only against the cheapest model it uses today. A router that falls back to a stronger model under load is a cost event. Model it before launch. A billing cycle is too late to discover it.

How often should this get reviewed

At minimum, monthly, and immediately after any provider price change or model upgrade. Treat AI feature cost the way a usage-based company treats cloud infrastructure spend. It is a line item with an owner. A number that only surfaces once a quarter in a finance review is reviewed too late.

A margin ceiling test takes an afternoon to run. It is cheaper than finding out from a finance review three months after launch. Aero can help you build one.

Sources

From the journal

Field Notes

Start a project

Make your product agent-ready

We design the products, websites and design systems your customers use and their AI agents act on.

  • Agentic product design
  • Design systems AI agents can read
  • Booking for Q1 2027