AI Agents
AI agent costs: what makes an agent cost more or less
A plain guide to the six things that decide what an AI agent costs to build and run, so you can ask better questions before you scope one.
AI agent costs come from six drivers: which model the agent uses, how many tool calls it makes, how much context it carries from step to step, how much human review it needs, how much integration work connects it to your systems, and how much monitoring it needs once live. Any quote for an agent, including AI agent pricing from a vendor, is a combination of these, so ask about each one.
plain answer
What drives AI agent costs
AI agent costs depend on how much work the agent does per task and how much care the task needs around it. There is no single price for an agent, because two agents that sound alike can differ a lot on every driver below.
Some drivers show up on the usage bill: the model, tool calls and context. Others are people and engineering time: human review, integration work and monitoring. A fair estimate covers both groups.
The sections below take each driver in turn and say what makes it larger or smaller.
explainer
Model choice
Language models are priced per model, and the price depends on how many tokens go in and how many come out. Models are priced differently from one another, and the pricing table shows that a more capable tier generally costs more per token than a smaller one, so the model you pick for each step is the first cost lever.
Many agents do not need one model for everything. A simple routing or extraction step may not need the model that writes the final answer. Splitting the work this way is a design choice, and it should be tested on your own cases before you rely on it. For a platform-specific look at routing work between models, see Cost-optimized model routing in Azure AI Foundry.
Output tokens count as well as input tokens. An agent that writes long answers or thinks at length before acting uses more output than one that returns short structured results.
explainer
Tool calls
An agent works by calling tools: searching, reading a record, running a calculation, writing to a system. Each call is another round trip to the model, and each round trip is billed.
Anthropic's documentation says tool use is priced on the input tokens sent to the model, including the tool definitions, the output tokens generated, and, for tools that Anthropic runs, an extra usage-based charge. The tool definitions, the tool calls and the tool results all add tokens.
So an agent with many tools, long tool descriptions or tools that return large results costs more per step. An agent that needs more steps to finish a task costs more than one that needs fewer. Clear tool descriptions and tools that return only what the agent needs both help.
explainer
Context length
Context is everything the model sees on a given step: the instructions, the conversation so far, the tool definitions and every tool result. In an agent loop it grows with each step, because earlier turns are carried forward.
Longer context means more tokens to process on every later step, so a long-running agent gets more expensive as it goes. Anthropic's documentation also notes that more context is not automatically better, because accuracy and recall can degrade as the token count grows.
Two things help. Prompt caching lets repeated parts of the input, such as fixed instructions, be reused at lower cost when they are read again from the cache. For long runs, clearing old tool results or compacting earlier turns keeps the context from growing without limit.
explainer
Human review
Human review is easy to leave out of an estimate. If a person checks every output before it goes anywhere, their time is part of the cost of the agent, and it grows with volume.
The amount of review depends on what happens when the agent is wrong. An agent that drafts an internal summary needs less review than one that changes a record or contacts a customer. Reviewing by exception, where only uncertain or high-impact cases go to a person, needs a clear rule for what counts as an exception and a way to test that rule.
This driver is general engineering reasoning, not a figure from a vendor page, and the right level has to be set for your process.
explainer
Integration work
An agent is only useful if it can reach the systems where the work happens. Connecting it to those systems is engineering effort, and it can be a large part of the build, depending on the systems involved. Our page on AI agent integration explains how an agent connects to those systems.
It grows with the number of systems, the quality of their interfaces, how the agent signs in, which records it may read or change, and how errors are handled. A system with a clean, documented interface is easier to connect than one that needs workarounds.
Teams should list the systems the agent must touch before asking for a scope, since each one adds work.
explainer
Monitoring and upkeep
Once an agent is live it needs watching. Prompts, models, data and the systems around it all change, and an agent that worked last month can drift.
Monitoring covers logging what the agent did and why, testing it again when something changes, and having a way to roll back. It costs engineering time and some tooling, and it keeps the other drivers from creeping up unnoticed. Kriv AI describes the method in its guide to an LLM evaluation framework.
how to
How to estimate your own AI agent costs
1. Describe one task the agent would do, start to finish, and how many times a day or month it would run.
2. List the steps and the tools each step needs. More steps and more tools mean more calls.
3. Decide which steps need a capable model and which could use a smaller one, then test that on real examples.
4. Estimate how much context builds up across the steps, and whether caching or clearing old results applies.
5. Decide what a person must review, and what happens when the agent is wrong.
6. List the systems to connect and who owns each.
7. Plan how you will monitor and re-test after launch.
8. Run a small pilot and read the real usage report. Anthropic's API reports the tokens each request used, which gives you measured numbers instead of guesses.
kriv approach
How Kriv AI scopes an agent
Kriv AI scopes agent work after a short discovery call. We go through the six drivers above for your process: the task, the systems involved, the review you need and how it will be monitored. Contact us to start the conversation.
limits
What this page does not cover
This page explains drivers. It does not give prices, rates or a total, because they depend on your task and on vendor pages that change. Check the current pricing page of the platform you use before you plan.
It also does not compare one approach with another on cost. For one platform's view of controlling agent costs, see Controlling Agentic AI Costs: n8n TCO for Regulated Mid-Market Teams.
Straight answers
Frequently asked questions about AI agent costs: what makes an agent cost more or less
What makes an AI agent cost more?
More capable models, more tool calls, longer context carried between steps, more human review, more systems to connect and more monitoring. Each driver is explained on this page.
What does AI agent pricing depend on?
It depends on the task, the model, how many steps and tool calls it takes, how much context it carries, the review it needs, the integration work and the monitoring after launch.
Do tool calls add to the cost of an agent?
Yes. Anthropic's documentation says tool definitions, tool calls and tool results add tokens, and some tools that Anthropic runs have an extra usage-based charge.
Why does a long-running agent get more expensive?
Because its context grows with each step and earlier turns are carried forward. Caching repeated input and clearing or compacting old results can limit this.
Is the model the biggest cost?
Not always. Integration work, human review and monitoring can matter as much as the model, depending on the task.
How do I get a cost for my agent?
Describe one task, list the steps, tools and systems, and decide the review you need. Kriv AI scopes agent work after a short discovery call.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call