Cost to Run an AI System: The Monthly Bill, Line by Line

The monthly bill for an AI system has two parts.

The first part is the model. That is the AI that reads and writes the text. You pay for it by the token, for what goes in and what comes out. A token is a small chunk of text, about a short word. So this part is plain math. Take the interactions, times the calls for each one, times the tokens per call, times the price.

A production system is one that is live, with real people using it. Some clients run them at 1,000+ AI interactions a day. For them, the model part typically lands between $200 and $1,200 a month.

The second part is what the model runs on. If the system talks, that means phone and voice minutes. If it answers from your files, it means a database and a search index. Then add hosting, background jobs, and the tracking that shows what each call cost. Our build library has estimates for six builds. The whole stack runs from $5 to $80 a month for a small drafting tool. At the top, a real-time voice coach runs $474 to $2,404.

That is the answer. The rest of this page is the detail, line by line, with the math.

What goes into the monthly bill?

Strip any live AI system down and the bill has the same five lines.

1. Model calls. Priced per million tokens. Input (what you send) and output (what comes back) each have a price.
2. Retrieval. This is how the system finds the right part of your files. You pay to turn them into a form it can search, called embeddings or vectors. You often pay again to rank what it finds, best match first.
3. Voice and messaging. Phone lines, speech-to-text, an AI voice, and text messages. You pay by the minute or by the message.
4. Data and hosting. The database, the search index, file storage, the servers and the job queues.
5. Monitoring. A trace of each model call with its cost, plus error alerts and logs.

Each build in our library has an estimated monthly tool cost, split by line. Six builds show how the mix shifts with the kind of system:

Read the third column. In the text systems, the model is the biggest line. So model choices move the total. In the voice systems, minutes are the biggest line. There, two things matter more than the model behind the voice. One is how long the calls run. The other is how many calls need a live voice at all.

These are estimates at the volume each build was made for. The usage lines grow with volume: model calls, minutes, messages. The hosting lines step up in tiers and barely move.

How to estimate the model bill

An interaction is one job done for one person, such as one question answered. The math is simple. Monthly model cost = interactions x calls per interaction x tokens per call x price per token. Here it is with numbers.

Hypothetical. The volumes are made up. The prices are samples, not a real price list.

• 1,000 interactions a day for 30 days: 30,000 a month.
• Each one makes 2 model calls: 60,000 calls a month.
• Each call sends 2,500 tokens in and gets 400 back.
• Input: 60,000 x 2,500 = 150 million tokens.
• Output: 60,000 x 400 = 24 million tokens.
• Sample price on the top model: $10 per million tokens in. $25 per million out.

Input: 150 x $10 = $1,500. Output: 24 x $25 = $600. With every call on the top model, the bill is $2,100 a month. I call it the lazy model bill. It is what you get when no one chose which model does which job.

Now pull two levers:

Going from $2,100 to $552 saves $1,548 a month. That is a 74% drop. At ten times the volume, the lazy bill is $21,000. The routed, cached one is $5,520. You save $15,480 a month, or $185,760 a year. The saving is a share of the bill. So it grows as your traffic grows.

Two details trip up most first guesses. First, count calls, not chats. One customer question can set off four calls. One sorts it, one searches, one answers and one checks. An agent makes more. (An agent is an AI that takes steps and uses tools on its own.) Second, count all that rides along as input. Think of the system's rules, its tool list and the file passages put in each prompt. All of them are input tokens. You pay for them on every call.

Why do voice systems cost more?

Because a voice system pays for time as well as words. A call needs a phone line. It needs speech-to-text that listens the whole time. It needs an AI voice that speaks, and a model that decides what to say. Our library puts real-time voice at about $0.16 a minute for the voice coach above. A thousand 10-minute sessions a month is 10,000 minutes. That is about $1,600, before any other line.

So the levers are not the same as for text. Keep calls short. Use a text channel for requests that don't need a voice. Put a small, cheap model behind the voice for the routine parts of a call. Save the costly one for the moments that need it.

The Six Levers that keep the bill down

I call them the Six Levers. Each one comes with a question you can put to any vendor.

Routing is the big one. It sends each job to the cheapest model that can do it well. Most of what an AI system does all day is dull. Sort a ticket. Pull a date from a PDF. Answer the easy question. A small model does that well, at a fraction of the price. In a routed system, the costly model is an exception you can name. In the lazy model bill, it is the default no one chose. In the worked example, the small model takes 60% of the calls. It costs about 13% of the routed bill: $126 of $966.

Caching is priced by the providers, so you can check the math. Prompt caching means the provider stores the part of your prompt that repeats. Then it costs less to reuse. Anthropic bills cached input at 0.1 times the base input price when it is read back. Writing it to the cache, which lasts five minutes, costs 1.25 times the base price. So a prompt that repeats the same long instructions pays full price once. After that, it pays a tenth.

One more lever sits outside the six, because it only fits work that can wait. OpenAI sells a Batch API. An API is the line your software uses to send work to a model. With Batch, jobs run within 24 hours at a 50% discount. That suits nightly reports and office data jobs. It never suits a customer on the line.

What does a runaway loop cost?

Here is a weekend. It is made up, at the sample prices above. On the top model, one call at 2,500 tokens in and 400 out costs $0.035.

At 6 a.m. on Saturday, a tool the agent calls starts timing out. The agent does what agents do. It tries again, every two seconds. No one is watching. By noon on Sunday, 30 hours later, it has made 54,000 calls. That is $1,890. It is most of a month's lazy bill, on a system nobody uses.

Now add a daily budget alarm set at $100. It would have paged a named person before 8 a.m. on Saturday. A circuit breaker is a hard cap that stops a loop. Set one at 20 calls per interaction, and it stops this loop at the twentieth retry. Twenty calls at $0.035 is $0.70.

Those two levers cost the least to build. They are also the easiest to leave out of a pilot. A pilot never has a weekend like that.

Where does run cost go in the ROI math?

On the cost side. ROI is return on investment: what a system earns back for what it costs. A system's payback is the value it makes in a year, divided by its total cost. That total is the build cost plus the first year's run cost. We hold builds to three to four times. How to calculate AI ROI has the full method.

In the worked case, the first year's run cost is $25,200 on the lazy bill. On the routed, cached one, it is $6,624. Same build, same value. The lazy bill adds $18,576 a year to the cost side. The system has to earn that back before it clears the bar.

At scale, that is not a rounding error. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027. It names three causes: rising costs, unclear business value, or weak risk controls.

Whose account should the bill run on?

Yours. Put the model, cloud and hosting accounts in your company's name. Then each line of the bill comes to you, where you can read it. A bundled monthly fee that includes the tokens hides the very line this page is about. One buyer told us their last vendor held the keys and charged extra on the tokens. Another could never get a straight answer to "what's your cost?"

Ask each vendor two questions. Who pays the provider? What is added on top? Our builds run on accounts the client opens on day 0. That answers both. Code and account ownership is its own subject. Who owns the code an AI vendor writes covers it.

Five mistakes that make the estimate wrong

1. Budgeting from the pilot. A pilot that runs ten times a day costs next to nothing. Price the volume you expect once it is live.
2. Counting chats, not calls. One chat often makes two calls or more.
3. Forgetting the passengers. The rules, the tool list and the passages it finds ride along on every call. They are all input tokens.
4. Pricing only the model. In a voice system, minutes are most of the bill.
5. Giving the bill no owner. An alarm that goes to a shared inbox is just a log entry.

When to optimize, and when not to yet

A prototype can be sloppy. That is the early test version, built to learn. You are still finding out what the system should do. So send it all to the best model, and read what comes back. That is a good use of a top model.

With no volume yet, routing and caching aren't worth the work. At 100 interactions a day, the worked case costs $210 a month on the lazy bill. Tuned, it costs $55.20. That saves $154.80. Spend the work on whether the system does its job. Routing waits for real traffic anyway. You can't route what you have not measured. The traces from early traffic show you which jobs are dull.

Alarms and breakers are the exception. All the other costs on this page grow with traffic. A stuck tool and an agent that retries make their own. So a system with a $55 bill can have that Saturday too.