How Much Do AI Agents Cost? Tokens, Time, Rate Limits

How much do AI agents cost? It depends on your plan, the model and the size of your project, and the only honest way to know is to watch your own numbers. This guide explains tokens, the context window and rate limits in plain words, and shows what ULU displays so you can see cost and time as they happen.

AI agents are paid for in usage: either API money or the usage allowance of your Claude plan.
A conversation with an AI has a size limit called the context window, and reading files fills it.
A team of agents uses more than a single chat, because each agent has its own conversation and its own reading to do.
When a Claude plan's usage limit is hit, ULU pauses the task, keeps the computer awake, shows when the limit resets and resumes.
ULU shows the running cost in USD, the time and the files for every step, so you can stop a runaway task and learn what a feature costs.

The short answer to "how much do AI agents cost"

There is no single price, and anyone who gives you one without knowing your project is guessing. The cost depends on three things: the plan or API key you use, the model the agent runs on, and how much work your project needs.

This article gives no dollar figures on purpose. Prices change and plans differ. What we can do is explain what you are paying for, and show you where to look.

ULU is a desktop app for Claude Code. You bring your own Claude plan or API key, the agents run in Claude Code, and ULU shows what they cost. It shows the running number in USD for every step, so you see your own figure and not an average from a brochure.

Tokens: the unit AI work is counted in

An AI model does not read words the way you do. It reads small pieces of text called tokens. A short word is often one token. A long or unusual word may be split into several. Code, with its symbols and odd names, splits into many.

You do not need to count tokens. You only need to know three things about them:

Everything the AI reads costs tokens: your request, the files it opens, the earlier messages.
Everything the AI writes costs tokens too: its answers, the code, the tests.
More work means more tokens, and more tokens means more usage on your plan or more API money.

The context window: a conversation has a size limit

Every conversation with an AI has a maximum size. That limit is the context window. It is the amount of text the AI can hold in mind at one time: your instructions, the files it has read, what it has said and what it has done so far.

Picture a desk. You can put a lot of papers on it, but not an unlimited amount. Each time the agent opens a file, a sheet lands on the desk. When the desk is nearly full, the agent has to clear space, and it may lose track of details it read earlier.

Why reading files fills it

Software projects are made of files, and agents learn a project by reading them. A big file takes a big part of the desk. An agent that explores your whole project before it starts may fill much of its window before it writes one line.

When the window is nearly full, the agent gets less sharp. It may forget an earlier decision or read the same file again.

ULU shows how full each agent's context is, before it becomes a problem. A bee whose memory is nearly full appears as Tired in the hive, so you can see it without reading a log. You can watch the whole hive in the Watch the hive work section of the site.

Why a team of agents uses more than one chat

A team of agents uses more usage than a single chat for the same idea. That is true, and it is better to say it plainly. Each agent starts its own conversation, reads what it needs and writes its own result. Six agents mean six readings.

What you get for it is separation of jobs. In a ULU team, a Product Manager writes down what to build, an Architect or Tech Lead decides how, a Designer shapes the screens, a Developer writes the code, QA tests it, and Security and Release finish the job. Each one works in a fresh window, focused on one task, instead of one overloaded conversation trying to hold everything.

More agents means more reading and more usage. In return, focused windows keep each agent sharp, and roles that check each other catch mistakes before they reach you.

We make no claim about how much you save or lose by working this way. It depends on the project. If you want to see how teams are built and what each role does, read AI agent teams explained, or look at the ready stacks in the teams section.

Rate limits: what they are and what happens when you hit one

A rate limit is a cap on how much you can use in a given period. Claude plans include a usage allowance. When you have used it up, the service tells your tools to wait until the allowance comes back.

What ULU does when the limit is hit

ULU treats a usage limit as a pause, not a failure. When your Claude plan's limit is reached, it does four things:

It pauses the task.
It keeps the computer awake.
It shows when the limit resets, so you know how long the wait is.
It resumes without losing the session, once the limit has reset.

You can leave the desk and come back. The work continues from where it stopped, not from the beginning.

Why seeing cost and time per step matters

A bill at the end of the month tells you nothing about which task caused it. A number per step tells you exactly where the usage went, while it is happening. That gives you two abilities.

You can stop a runaway task

Sometimes an agent loops, reads far more than it should or keeps retrying something that does not work. If you only see the result later, the usage is already spent. If you see the cost climbing step by step, you can stop it. You can also continue later, even days later.

You can learn what a feature costs

After a few builds you have real numbers from your own work. You learn what a login page cost, what a change to the checkout cost and how long each took. That is far more useful than a general price, and it lets you plan the next request.

The numbers are shown for each stage, each build step and each agent. See them in the tracking section and the four steps.

What you see per step

The four readings ULU shows for each step, and what each one tells you
Reading	What it shows	What it tells you
Cost	The running cost in USD, per step and in total	Where your usage is going, and whether a step is costing more than you expected
Time	Running time now, total time when the step is finished	Whether an agent is making progress or stuck, and how long a feature takes
Files	Files written, tools used, agents started	What the agent actually did, so you can match the cost to real work
Context fullness	How full each agent's memory is	Whether an agent is close to its limit and about to lose track of earlier details

What the brain does for cost

Much of the usage in a project goes to reading. Every new agent has to learn the project again, and every new task starts with the same exploring. That repeated reading is a cost you pay each time.

ULU keeps a project-level memory, called the Brain. It records how files change together, which tests cover what and what the team learned. The idea is simple: a team that remembers does not need to read everything again.

We do not claim a number for how much this saves, because it depends on the project and we have no figure to give you. What we can say is where the saving comes from: less re-reading means fewer tokens spent on learning what the team already knows. You can check this yourself, by comparing the cost of similar steps over time.

Schedules: running work in quiet hours

You can run a team on a schedule, for example at night. There are two good reasons to do it.

The computer is idle anyway, so the work does not get in your way.
If the work hits a usage limit, ULU pauses and resumes, so it does not need you awake to carry on.

Scheduled teams work from an inbox of tickets that you approve. The same cost and time readings are there when you come back, so you can check what the night cost.

Where to read more

If you are new to the whole idea, Claude Code for people who don't code explains the engine underneath. If you want to know why a team of bees is organised like a small company, what an AI software factory is covers the full picture. And if you already know your tools, a ready AI team for your tech stack shows the ten starting points.

Questions people ask

How much do AI agents cost?

It depends on your Claude plan or API key, the model the agents use and the size of the work. We do not quote a price, because we cannot know your project. ULU shows the running cost in USD for every step, so you see your own number as the work happens.

What does "context window full" mean?

The context window is the maximum amount of text a conversation can hold: your instructions, the files read and everything said so far. When it is nearly full, the agent can lose track of earlier details. ULU shows how full each agent's window is, so you can act before it becomes a problem.

What happens when I hit the Claude rate limit?

ULU pauses the task, keeps the computer awake, shows when the limit resets and resumes without losing the session. You do not have to restart the work. You can wait at the desk or leave and come back.

Does a team of agents cost more than one chat?

Usually it uses more, because each agent has its own conversation and does its own reading. In return, each agent has one job and a focused window, and the roles check one another. Whether it is worth it depends on how big and how important your project is.

Try it

Install ULU and run your first project, which is free. Watch the cost, time and context for each step, and you will know what your own work costs after one build. Install ULU.