LLM API cost calculator (tokens)

How to calculate the monthly cost of an AI API

The calculation starts with three numbers you need to know before you even look at the price list: how many requests you have per day, how many tokens on average go into each request, and how many tokens the model returns. Most people make the mistake of estimating only the length of the question and forgetting the context sent with each question: system instructions, previous messages in the conversation, extracted parts of documents. In conversational applications, context often makes up 70 to 90 percent of input tokens.

The procedure:

1. Take ten real requests from production or tests and copy them into a text file.
2. Measure input tokens with the provider's tool (tokenizer) or a rough estimate: number of characters divided by 3 for Serbian.
3. Measure output tokens the same way, but only for the responses the model actually returns.
4. Multiply the averages by the number of requests per day, then by 30 for the month.
5. Apply the prices per million tokens for input and output separately and add them up.

Example: 2,000 requests per day, 1,500 input and 300 output tokens. That is 3 million input and 600 thousand output tokens per day, or 90 million input and 18 million output tokens per month. If input is 1 currency unit per million tokens and output is 4, the monthly cost is 90 + 72 = 162 units. This kind of itemized calculation immediately shows where to cut.

What is a token and why Serbian text costs more

A token is not a word or a letter, but a piece of text that the model processes as a single unit. English words are often split into one token, while longer Serbian words with suffixes are split into two or three. That is why the same content in Serbian takes up more tokens than in English, and the difference is greatest with case forms, compound words and words with diacritics.

The practical consequence: if you translate a prompt from English and do not adapt it, you pay more than you think. Short instructions in Serbian are fine, but long templates and product descriptions are worth writing as concisely as possible. The same applies to JSON output: every key, quotation mark and bracket counts.

Tip for advanced users: instead of guessing, create a small test set of 20 to 30 examples and measure tokens programmatically before each call. Those measurements later serve as the basis for your budget and for comparing models, because the difference between two models often only shows up on real data, not on a demo prompt.

Most common cost estimation mistakes

• Counting only the last question. The conversation history that is resent with every request gets forgotten.
• Ignoring output tokens. People enter the input price and think that is the whole cost, but output is usually several times more expensive.
• Averages without tails. Rare requests with a huge document can outweigh hundreds of ordinary ones. Check the largest requests too, not just the average.
• Outdated prices. Providers change price lists and introduce new model classes, so a calculation from a month ago can be wrong.
• Large numbers of tiny requests. With very short responses the cost per request is small, but 50,000 requests per day still adds up to a serious bill.

How to reduce cost without losing quality

First, check whether every request really has to go to the strongest model. Routine tasks such as classification, extracting fields from text and short summaries are often handled by a smaller model with a similar result. Second, shorten the context: instead of the whole document, send only the relevant passages, and limit the conversation history to the last few messages. Third, cache answers to repeated questions and set a hard limit on the number of output tokens, because that prevents rare but expensive "novels" in the response.

If cost is growing faster than the number of users, the problem is usually in the call architecture, not in the model price. EVI Solutions can help when you need to introduce spending tracking per user and project and set budgets at the application level.