Blog ·

How to price an AI feature when your model costs less per call

Agencies do not agree on how to price AI work. Productive surveyed 174 agencies in March 2026 and found no shared pricing model. Some raised prices, some cut them, and a growing group was still figuring it out. Among agencies that reported several positive results from AI, 52% held or raised prices (Productive). Promethean Research put average agency net margin at 13% in 2025 (Promethean), so how you price a feature decides a large share of what you keep.

When the model behind a feature gets cheaper, someone keeps the difference. Below are the numbers for one feature and three ways to price it.

The feature

A support-ticket router for an e-commerce client: 100,000 tickets a month, each about 800 input tokens, with a label of about 20 output tokens as the answer.

The model cost

At list prices (Anthropic, OpenAI, Google):

Model Per 1,000 calls Per month at 100,000 calls
Claude Sonnet 5.5 $1.80 $180
Gemini 3.8 Flash $0.68 $68 (doubles from 1 January 2027)
Claude Haiku 5.5 $0.09 $9
GPT-6 Luna $0.09 $9
Fine-tuned GLM 5.3 Flash on Honed, measured $0.21 $21

Two things follow from this table.

First, at this volume the model bill is a few hundred dollars a month at most, which is small next to the hours you spend building and maintaining the feature. Pricing the feature as "API cost plus markup" leaves most of your value off the invoice.

Second, the cheapest prompted models and a fine-tune are close on cost. Haiku is cheaper per call than our measured fine-tune, which ran on the largest base. The lowest-cost bases came in at $0.05 to $0.12 per 1,000 calls on our other tasks. The difference that shows up in the client's business is accuracy. On our support-routing run, the fine-tuned model got 100% of held-out tickets right and Claude Haiku 5.5 got 90.5%. At 100,000 tickets a month, that is about 9,500 more tickets reaching the right team without a person re-routing them, and that is the number to price on.

Three pricing models

1. Build fee plus a fixed monthly retainer. You charge for the build and the first eval, then a flat fee that covers hosting, monitoring, a monthly report and retraining when the data changes. The model cost sits inside the retainer. When a cheaper model arrives, your margin grows and the client's invoice does not move. This is the simplest to sell and the easiest to forecast.

2. Per-outcome pricing. You charge per ticket routed, per invoice extracted or per lead scored. Set the price against what the client pays now for the same work, whether that is a person's time or a more expensive model. The client sees a cost that scales with use. You carry the risk that volume drops.

3. Pass-through plus a management fee. The client pays model costs at cost, on their own account or with a stated cap, and you charge a monthly fee for the work. It is transparent and it suits clients who already have their own API contracts. It leaves the savings from a cheaper model with the client.

Most agencies already mix models; fewer than 8% use only one (Promethean). We suggest option 1 for the first year, then option 2 once volume is stable.

What to put in front of the client

Give them one page with four lines:

  1. The task and the monthly volume.
  2. The result table: accuracy, cost per 1,000 calls and median latency for the fine-tune and for the model they would use otherwise.
  3. What the accuracy difference means in their units: tickets, invoices, hours.
  4. Your price, and what it includes: retraining, monitoring and the test set every new version runs on.

The result table, built from their own data, is the part a competitor quoting "GPT plus a prompt" cannot show.

Sources