Buyer guide
What will an AI feature cost each month to run once it is built?
In short
AI providers bill by usage, mainly by the amount of text going in and coming out, so the monthly cost follows how many people use the feature and how much text each request carries. Ask for an estimate with its assumptions, choose the smallest model that does the job, and set spend limits and alerts.
Who is this guide for?
Owners planning to add a chatbot, assistant or text feature to their app who want to understand what the ongoing bill depends on, apart from the one-time build.
Build cost and running cost are different
Building an AI feature is a one-time piece of work. Running it is a monthly bill from the model provider, plus your normal hosting. The second figure is the one that surprises owners, because it grows with use.
Ask any developer for both numbers separately, and for the assumptions behind the running estimate.
How model providers bill
Anthropic's pricing page lists separate rates for input tokens, output tokens and cached input, and says that processing in batches and the way cached text is used affect what you pay. OpenAI's pricing page likewise bills text models by token, with separate rates for input, output and cached input, and different rates by processing tier and by the length of the input.
A token is a small piece of text. In practice, a request that sends a long document, or a long conversation history, costs more than a short question, and a long answer costs more than a short one.
Built-in tools can add charges of their own. OpenAI notes that tool use is billed at the model's token rates and may add charges per call, per session or for storage.
What drives your monthly total
Four things move the bill: how many requests you get, how much text each request sends, how long the answers are, and which model you use. A larger model costs more per token than a small one.
Design choices matter as much as traffic. Sending the whole conversation each time, or a whole knowledge base with every question, multiplies the text. Retrieving only the relevant passages and keeping answers short keeps it down.
Limits and alerts
OpenAI describes spend limits that are set separately from usage limits. A spend alert sends a notification without stopping traffic, while a hard spend limit makes requests fail once the cap is reached. Rate limits are measured in requests and tokens per minute or day.
Choose deliberately. An alert warns you, and a hard limit protects the budget but switches the feature off. Many owners use an alert at a lower figure and a hard limit above it.
How to compare two estimates
A fair estimate states the model, the average text per request, the number of requests it assumes per month and what happens if usage doubles. If it is only a single number, ask for the assumptions.
Provider prices change, so check the current pricing page rather than a figure quoted from memory, and ask how hard it would be to switch to another provider.
How I give an estimate
I would run a short test on real examples of your content, measure the text per request, and give you a running estimate with its assumptions. The build is quoted after a short discovery call, against a written scope.
How do the options compare?
| Cost driver | Why it matters | How to control it |
|---|---|---|
| Number of requests | Each request is billed | Rate limits, caching of repeated answers |
| Text sent per request | Input tokens are billed | Send only relevant passages, trim history |
| Length of answers | Output tokens are billed | Ask for short answers, set a maximum |
| Choice of model | Larger models cost more per token | Use the smallest that passes your tests |
| Tools and extra services | May add per-call or storage charges | Read the tool pricing before enabling them |
What should I check before I sign?
- The estimate separates the build from the monthly running cost.
- It states the model, the text per request and the monthly volume assumed.
- You know what happens to the bill if usage doubles.
- A spend alert and a spend limit are set in your own provider account.
- The provider account and billing are in your company's name.
- Answers and requests are kept as short as the feature allows.
- You have checked the provider's current pricing page.
- You know how hard it would be to change provider or model.
How pricing works
I quote in writing after a short discovery call, against a written scope. Building an app from scratch? The app cost calculator gives a first range in two minutes.
What else do people ask?
Is the monthly cost fixed?
No. Providers bill by usage, so it follows how much the feature is used and how much text each request carries.
Can I put a cap on the bill?
Providers offer spend alerts and limits. An alert only notifies you, while a hard limit makes requests fail once reached, so decide which you want.
Will a bigger model always give better results?
Not always. Test the smaller model on your own examples first, since it costs less per token.
Do I need my own account with the provider?
It is wise. Keep the account and billing in your company's name so you can see usage and change the setup later.
Sources
How was this page made?
I drafted this page with AI assistance from the official documents listed under Sources. Found a mistake? Tell me.
How can we work this way together?
- AI Features RetainerIn the AI features retainer I look after the LLM features in your product month by month: I tune prompts and retrieval, watch quality and cost, add guardrails and fallbacks and move you between model providers when needed.
- Cloud Cost ReviewIn the cloud cost plan I read your AWS or cloud bill against what you actually run, explain it in writing and then, month by month, remove waste, right-size servers and storage and test every change on the live system.