How to price AI work for clients when model costs move every month

Why AI running costs break normal agency pricing

A website costs roughly the same to host whether ten people visit or ten thousand. An AI feature doesn’t: every request is billed by the provider per token, in and out. The bill moves with three things you don’t fully control — how much the client’s users use it, how long the prompts get as features grow, and which model each call runs on. Price the running cost as if it were hosting and a successful launch becomes your loss.

Do the token math first

Before choosing a pricing model, estimate the monthly cost from three numbers: requests per month, average input tokens, average output tokens. An illustrative support assistant handling 200,000 requests a month, with 1,500 input and 300 output tokens each, at the list prices in our rate table:

Same traffic, a 16× difference — and most real systems are a mix: some calls need the premium model, many don’t. Which model each call runs on is the single biggest lever on your margin, so know the mix before you quote.

Model 1 — A retainer that includes usage, with a cap

One monthly fee covers build support and AI usage up to an agreed ceiling. Clients like the predictability; you carry the usage risk. It works when you can enforce the ceiling technically, not just in the contract.

• Quote the retainer on your estimate at expected volume, and set the cap at the volume where your margin goes to zero.
• Write an overage clause: above the cap, usage is billed at cost plus a percentage, or the feature pauses.
• Enforce the cap at the request level. A spend limit that only shows up in next month’s provider invoice isn’t a cap.

Model 2 — Cost-plus: pass usage through with a markup

You bill a build or support fee, and AI usage separately at cost plus 10–30%. You carry almost no usage risk and earn on growth. The catch is operational: you need per-client metering you can show the client, because “our OpenAI bill was $4,100, your share is $1,260” invites an argument without line items. Passing LLM costs to clients covers the ways to set this up.

Model 3 — Value-based: a share of what you save or earn

For optimization work — taking an existing AI feature and making it cheaper — price on the outcome: a fixed fee plus a share of the measured savings. It sells itself when the baseline is clear, and fails when it isn’t. The savings figure has to be measured per request against what the call would have cost, not projected from a sample, or neither side will trust it.

Protecting your margin, whichever model you pick

• Meter per client from day one. Separate projects and keys per client, so every request is attributable.
• Cap per client. A hard monthly ceiling that refuses requests before the provider is called.
• Route the easy calls down. Short, simple, low-temperature calls rarely need the flagship model.
• Cache exact repeats. Support and internal assistants answer the same question all day; an exact cache hit costs nothing.
• Report monthly. A client who sees spend, savings and trend every month renews; one who sees a surprise invoice shops around.

Where ModelGate fits

ModelGate is a gateway between your clients’ apps and the model providers. It prices every request to the token per client, enforces a monthly spend cap per client, caches exact repeats and routes simple calls to a cheaper model when you switch that on, and produces a per-client savings report without its own fee on it — so any of the three pricing models above has the numbers behind it. See how it works for agencies.

Keep reading

• Passing LLM API costs through to clients: three models that work — The client's own provider account, rebilling with a markup, or a bundled fee — and what each needs.
• What to put in a monthly AI cost report for a client — The seven numbers a client actually reads, and how to get them without a spreadsheet.
• ModelGate for agencies — per-client ledgers, caps, reports and billing.