How to calculate the cost of an AI project based on tokens

Before building an AI feature, there's a question your client, CFO, or your own common sense will ask: how much will this cost per month? Answering vaguely is the fastest way to an unpleasant surprise on the bill. The good news is that, with the right knowledge, you can estimate the cost of an AI project with considerable accuracy before writing a single line of code.

This article shows the step-by-step method for calculating the cost of a token-based project.

The starting point: the formula

The entire cost of an AI API boils down to this multiplication per call:

cost per call = (input tokens ÷ 1 million × input price) + (output tokens ÷ 1 million × output price)

From here, just multiply by the call volume. The entire estimate rests on four variables: input tokens per call, output tokens per call, number of calls, and the price of the chosen model.

Step 1: estimate tokens per call

First, understand how much text goes into and comes out of a typical call. Use the rule of thumb that one token equals about 0.75 words, or that one million tokens are about 750,000 words.

Define a typical case. For example, in a support chatbot, each question might involve about 500 input tokens (the question plus system instructions) and generate 300 output tokens. In a report generator, the input might be 2,000 tokens and the output 500. Be realistic and use an average case, not the best-case scenario.

Step 2: estimate call volume

How many calls per day or per month do you expect? Link the number to the actual product usage. One thousand users making ten requests per day means 300,000 calls per month. This is where tiny individual costs turn into amounts that matter.

Step 3: choose the model and price

With estimated tokens per call and volume, choose the model and use its price per million tokens. Remember that output almost always costs five times more than input, and the choice of model is the factor that most alters the final result.

Step 4: do the math

Put it all together. A complete example:

• Support chatbot, 500 input tokens and 300 output tokens per call.
• Volume of 300,000 calls per month.
• Model at $1 input and $5 output per million.

Input per month: 300,000 × 500 = 150 million tokens, at $1 per million, equals $150.

Output per month: 300,000 × 300 = 90 million tokens, at $5 per million, equals $450.

Estimated total: about $600 per month.

Notice how the output, even with fewer tokens, dominates the calculation because it's more expensive.

Step 5: adjust for reality

The base estimate is your starting point, not the final number. Adjust it for the factors that reality brings:

• Safety margin. Add 20 to 30 percent for usage variations and longer-than-expected responses.
• Growth. If the product grows, the cost grows proportionally. Project future volume.
• Savings. If you plan to use caching or batch processing, apply the corresponding discounts, which can cut a large part of the cost for suitable workloads.
• Spikes. Long conversations and extended reasoning functionalities consume more than the average. Identify them.

A tip for proposals and budgets

When presenting this to a client or leadership, always provide a range, not a single number. Something like between 500 and 800 euros per month at this volume conveys rigor and protects you against natural usage variation. Add the volume premise on which the estimate is based, so that any change in usage can be easily re-evaluated.

Conclusion

Calculating the cost of an AI project is not guesswork; it's multiplying four known variables: input tokens, output tokens, call volume, and model price. With a realistic estimate per call and expected volume, you'll arrive at a defensible number before building anything.

Mastering this calculation is what allows you to propose AI projects with confidence, size the architecture for the right budget, and prevent the bill from becoming the most memorable part of the project.