What Does Running AI Cost Per Month?

Running AI is billed in tokens, which are pieces of text, not per user or per month. The running cost of a single automation is typically tens of dollars a month rather than thousands, and it can be calculated in advance from three numbers: how often the automation runs, how much text it reads, and how much it writes. The biggest cost risk is not a high price. It is that nobody owns the number.

The price of automation is usually discussed as the cost of building it. That makes sense, because the build is what gets quoted. After go-live, though, a second number remains, and it repeats every month for as long as the automation is in use. This article covers how that number is built and how it is kept in check.

How is the running cost of AI built?

The running cost comes from tokens. A token is a piece of text, roughly 0.75 words in English, and every AI call is billed on two numbers: how many tokens the model read (input) and how many it wrote (output). Output typically costs five times as much as input.

Model choice is the third multiplier. The same work can be done with a light or a capable model, and the difference is fivefold:

Model	Input / million tokens	Output / million tokens
Claude Haiku 4.5	1 dollar	5 dollars
Claude Sonnet 5	2 dollars	10 dollars
Claude Opus 5	5 dollars	25 dollars

For practical scale: according to Anthropic's own worked example, handling 10,000 customer support tickets consumes roughly 3,700 tokens per conversation, and the whole batch costs about 37 dollars on the lightest model. A thousand tickets therefore costs under four dollars. That is the right order of magnitude, and most people guess it far too high.

These figures are about text, meaning reading text and producing it. The same models also handle images, documents, and speech, and the principle there is identical: the content is turned into tokens and the tokens are billed. Only the volumes are in a different league. An image is split into 28 by 28 pixel patches, so an ordinary 1,000 by 1,000 pixel image is about 1,300 tokens and a 4K image as much as 4,800. A single 4K image therefore consumes more than the entire support conversation above. If an automation handles scanned documents, photos, or audio, count the tokens on that basis rather than on the amount of text.

How much does a typical automation consume per month?

Take an automation that reviews the deals in your sales pipeline once a week and writes a short risk assessment for each one. Assume 200 deals, 8,000 tokens of data and instructions read per deal, and 1,200 tokens of written assessment. That is 800 runs a month.

Model	Cost per run	Cost per month
Claude Haiku 4.5	0.014 dollars	about 11 dollars
Claude Sonnet 5	0.028 dollars	about 22 dollars
Claude Opus 5	0.070 dollars	about 56 dollars

The same logic applies to any automation: number of runs times tokens per run times the model price. If those three numbers are known, the monthly cost can be calculated before a line of code is written.

Note that this is a different number from the cost of building it. When you compare an off-the-shelf tool with a custom implementation, the total cost of both is made up of the build, the running cost, and the upkeep. We covered that comparison separately in when a custom AI solution beats an off-the-shelf tool, where the frame is three-year total cost.

What gets billed on top of the token cost?

Tokens are only one line on the invoice. Off-the-shelf agent and automation platforms typically also bill for runs and for the tools used, and those line items can exceed the cost of the model itself.

Typical charges on top of tokens:

• Runs or activations. Many platforms price by how many times a workflow starts, not by how many tokens it consumes.
• Tools and integrations. Every connected system can carry its own fee, and on some platforms integrations are billed separately from the licence.
• Server time. Running code or keeping an agent session alive is often billed by time. In Anthropic's pricing, agent session runtime is 0.08 dollars per hour and code execution 0.05 dollars per hour per container once the month's 1,550 free hours are used.
• Searches and other external calls. Web search is the typical example: it costs 10 dollars per thousand searches on top of tokens.

The practical consequence is that two quotes cannot be compared on token price alone. Always ask what is billed on top of tokens and in what unit. In a custom build the line items are usually more transparent, because you pay the model and infrastructure cost directly without a platform margin, but the infrastructure is then your responsibility too.

Why can the cost multiply without anyone noticing?

The running cost rarely rises because of a price increase. It rises because one of the three multipliers changes quietly, and nothing shows it before the invoice.

The four most common causes:

• The amount of text read grows. An automation that started by reading the basic details of a deal is reading the entire conversation history a year later. Tokens triple while the functionality looks identical.
• An agent takes many loops instead of one. An AI agent that plans its own steps easily makes ten calls rather than one while completing the same task. The difference from traditional automation is the same here as it is in behaviour, and we covered it in more detail in what an AI agent is and how it differs from automation.
• A model is switched to a more capable one during a test and never switched back. A fivefold price stays in place because the experiment was never unwound.
• The number of runs grows with the business. Twice the customers means twice the invoice. That is a good problem, but it needs to be known in advance.

One technical detail is worth knowing: the newest models use a new tokenizer that produces roughly 30% more tokens for the same text than the previous one. A model upgrade can therefore raise the bill even when the price per token stays the same.

How is the running cost kept under control?

The cost is kept under control with four moves, and they are worth making in this order. The first three cost you nothing in quality.

1. Turn on caching. If every run reads the same instructions and the same background material, that content should be cached. Cached text costs 10% of the normal input price. In our example automation, where 6,000 of the 8,000 tokens are identical every time, this drops the monthly cost from 22 dollars to about 14. Caching pays for itself after the first read.
2. Move non-urgent work to batch. Work whose result is needed tomorrow rather than in five seconds can be run in batches. Batch processing is 50% cheaper on both input and output, and the discount stacks with caching. A weekly report that runs overnight does not need real-time handling.
3. Strip unnecessary text from the input. Most automations that cost too much are reading data they never use. Removing the fields the model does not need lowers the cost in direct proportion. This is the same exercise as scoping the process, so do it at the same time. A good basis for it is scoring processes to pick the first automation target, where cost is a natural fourth measure.
4. Choose the model to fit the task, not just to be safe. Classification, extraction, and summarization work fine on the lightest model. The most capable model is worth reserving for work where reasoning genuinely decides the outcome. This is the only one of the four where the saving can cost you quality, so do it only after the first three are done.

This discipline has a name, tokenomics, and in practice it means the same thing as FinOps does for cloud spend: consumption is measured, given a budget, and deviations are surfaced automatically. An SME does not need a dedicated tool for this, but it does need a named owner and a monthly number.

If you want a rough order of magnitude for your own situation first, the automation potential self-assessment tells you which processes to look at first.

What should you ask a vendor before signing?

Before a build contract, ask five questions whose answers are numbers or a yes and a no, not estimates.

• What is the estimated monthly running cost at current volume, and which numbers is the estimate based on?
• What is billed on top of tokens, and in what unit: runs, tools, integrations, server time?
• What happens if volume doubles?
• Is caching in use, and what share of the input is cacheable?
• Which model is used at which step, and who decides when it changes?
• How is consumption reported to us, and who notices a deviation?

The last question is the most important of the five. A cost nobody monitors is the same class of risk as an unmonitored AI output. Model choice also affects where your data travels, so this conversation is worth having at the same time as the AI security checklist before adoption.

Summary

Running AI costs less than most people fear. The problem is not the price but that the number is never budgeted and never watched. The monthly cost is the product of three numbers: runs, tokens per run, and the model price. A single automation typically lands in the tens of dollars per month. On an off-the-shelf platform, run, tool, and integration fees come on top, and in an automation that handles images or documents the token counts are considerably higher than for text.

Keeping it under control does not require a new tool. It requires caching, batch processing for non-urgent work, a stripped-down input, and a model that fits the task, plus one person who looks at the number every month. These are worth deciding during the build, because afterwards they become change work.

---

Empirica Finland specializes in AI solutions for B2B environments and is a Claude Partner Network member and a Microsoft partner.

← Back to homepage