Tokens, prices and the formula
A token is a piece of a word. For English text a common rule of thumb is about ¾ of a word per token, so 1,000 words are roughly 1,300 tokens, but it varies by model and language. Provider dashboards or a tokenizer tool give exact counts, and a test run on real prompts gives the averages you need.
Prices are quoted per 1 million tokens, and output usually costs several times more than input. The cost per request is =D2*B2/1000000+E2*C2/1000000 with input tokens in D, output tokens in E and the two prices in B and C. Because one request costs a fraction of a cent, analysts often quote the cost per 1,000 requests: the support reply in the example costs $0.004695 per request, or $4.70 per 1,000.
From cost per request to the monthly bill
The monthly cost of a use case is cost per request × requests per month, and the bill is the sum over all use cases: =SUM(H2:H5) = $473.40, or about $5,681 a year. The busiest use case is not always the most expensive one: 1,200 contract checks on the large model ($117.90) cost five times as much as 25,000 invoice extractions on the small one ($23.40).
For a forecast, model the drivers separately: users × requests per user per day × working days × tokens per request. Then you can see what happens when adoption doubles or prompts grow.
Worked example: monthly LLM cost for four use cases
Prices per 1 million tokens, average tokens per request and requests per month are yellow inputs. Cost per request, cost per month and the totals are formulas.
| A | B | C | D | E | F | G | H | |
|---|---|---|---|---|---|---|---|---|
| 1 | Use case | Input $ / 1M | Output $ / 1M | Input tokens | Output tokens | Requests / month | $ per request | $ per month |
| 2 | Support reply (Model M) | 1.50 | 6.00 | 1,850 | 320 | 60,000 | 0.0047 | 281.70 |
| 3 | Invoice extraction (Model S) | 0.30 | 1.20 | 2,400 | 180 | 25,000 | 0.0009 | 23.40 |
| 4 | Contract check (Model L) | 7.50 | 30.00 | 9,500 | 900 | 1,200 | 0.0983 | 117.90 |
| 5 | Sales call summary (Model M) | 1.50 | 6.00 | 6,200 | 550 | 4,000 | 0.0126 | 50.40 |
| 6 | ||||||||
| 7 | Total per month | 473.40 | ||||||
| 8 | Per 1,000 support replies (G2 × 1,000) | 4.70 | ||||||
| 9 | Output share of the bill | 35.1% | ||||||
| 10 | Total per year | 5,681 | ||||||
- G2
=D2*B2/1000000+E2*C2/1000000Cost per request: input plus output - H2
=G2*F2Cost per month for the use case - H7
=SUM(H2:H5)Monthly bill - H8
=G2*1000Cost per 1,000 support replies - H9
=SUMPRODUCT(E2:E5,C2:C5,F2:F5)/1000000/H7Share of the bill from output tokens
Illustrative price tiers (Model S, M, L), not any vendor's price list. Check your provider's current prices and date them in your model.
What drives the bill
Model choice dominates: the large tier costs 25 times the small one per token. A contract check costs $0.098, more than 100 invoice extractions. Output length comes next, because output tokens are priced higher: 35% of this bill is output although output is only about 12% of the tokens. Prompt size is the hidden driver: long system prompts, retrieved documents and chat history are sent with every request.
Real usage is usually higher than a naive estimate: retries, multi-step agents that call the model several times per task and growing conversation history all add tokens. A buffer of 20% to 50% on top of the test-run averages is a sensible starting point until you have a month of real data.
Caching and batching savings
Many providers discount input tokens that repeat across requests, such as a fixed system prompt, through prompt caching. Discounts and rules vary by provider, and some charge extra to write the cache. As an illustration: if 1,500 of the 1,850 input tokens of the support reply are cached at 10% of the normal input price, its cost falls from $0.004695 to $0.00267 per request, 43% less.
Batch processing, sending non-urgent requests to be processed later, is often offered at a discount too. It suits back-office jobs such as invoice extraction or review summaries, not live chat.
Choosing a model on cost and quality
The cheapest model is only cheaper if it is good enough. Score the candidates on a test set for quality, latency and cost, set a minimum quality for the task, and pick the cheapest model that passes. Then put the cost into the business case next to the time saved and the cost of review, so the AI budget is compared with what it returns.
Common mistakes
- Forgetting to divide by 1,000,000. Prices are per million tokens, so a missing division makes the bill a million times too high.
- Using one price for input and output. Output tokens usually cost several times more.
- Counting only the user's message. System prompts, retrieved documents and chat history are input tokens too, on every request.
- Ignoring retries and agent steps. One task can mean several model calls. Measure calls per task in a test run.
- Comparing models on price alone. A cheaper model that needs more human corrections can cost more overall.
- Undated prices. Provider prices change often. Record the date and source of every price in the sheet.
Frequently asked questions
How do you calculate the cost of an LLM API call?
Multiply the input tokens by the input price per million and divide by 1,000,000, do the same for the output tokens with the output price, and add both. Multiply by the requests per month for the monthly cost.
How many tokens is 1,000 words?
For English text roughly 1,300 tokens, using the rule of thumb of about ¾ of a word per token. Other languages and code often need more tokens per word, so measure with your own texts.
Why are output tokens more expensive?
Generating text takes more computation per token than reading it, so providers usually price output several times higher than input. That is why long answers raise the bill quickly.
How can I reduce LLM costs?
Use the smallest model that meets the quality bar, shorten prompts and answers, cache repeated prompt parts, batch requests that are not urgent and track cost per use case so you see where the money goes.