An AI agent has two costs: building it once, and running it on every task. The build is a project. The running cost repeats with every lead, call or ticket the agent handles, so it decides whether the agent pays for itself. This page covers the running cost: what it is made of, what makes it grow, and how to measure it before you commit.
Provider prices change often, so we don't quote them here. Check each provider's current price page, and measure your own cost per task on sample runs.
What you pay for on every task
- Model calls. Most providers charge per token, for the text the model reads and the text it writes. An agent that reads a long history or a large document on every step pays for that text every time.
- Tool calls. Search, database lookups and third-party APIs the agent uses can carry their own fees.
- Voice, for voice agents. Speech-to-text, text-to-speech and the phone minutes themselves, usually billed per minute.
- Retrieval. Embeddings and a vector database when the agent answers from your documents.
- Hosting and monitoring. The servers the agent runs on, plus logging and tracing so you can see what it did.
- People. Time spent reviewing the agent's work, especially in the early weeks, when it can be a large share of the total.
What makes the cost grow
- Context size. Sending the whole conversation or a whole document on every step multiplies the tokens.
- Steps per task. An agent that plans, retries and checks itself makes several model calls for one task.
- Model choice. Larger models cost more per token. Many steps work well on a smaller model.
- Call length, for voice agents: every extra minute is paid for in speech, voice and telephony.
- Loops and bugs. A task that repeats by mistake keeps costing until someone notices.
How to keep it under control
- Measure cost per task on sample runs before launch, with your real inputs, not a demo.
- Route each step to the smallest model that does it well, and keep the large model for the steps that need it.
- Make the model a setting, so you can switch provider when prices or quality change. Read: make a model switch cheap
- Set budgets and alerts per agent and per key, so a loop or a leaked key is caught early. Read: make a stolen AI key worth less
- Log cost per customer or per task, so you know which work is worth automating.
