LLM Cost Estimator
Estimate the cost per request and the total for your request volume, from token counts and per-million-token pricing — input and output.
Prices are estimates you control — confirm current rates with your provider. Caching, batch discounts, and thinking/tool tokens can change the real figure.
What this tool does
The LLM Cost Estimator turns two prices and three volume numbers into a dollar figure. You enter the input and output prices in dollars per million tokens, the average input and output tokens in one request, and the number of requests. It returns the cost of one request and the total for all of them.
It answers the early budgeting question: roughly what will this feature cost at the volume we expect, and how much does that change if the prompt grows or the model is pricier? It has no vendor price list, so the numbers you type are the only numbers it uses.
How it works
The tool prices each request as input tokens times the input rate plus output tokens times the output rate, then multiplies by the number of requests. Rates are per million tokens, so token counts are divided by 1,000,000 first. It calculates as you type, in your browser, with no network request.
- Cost per request
- perReq = (inTok / 1,000,000) x inP + (outTok / 1,000,000) x outP
- Total
- total = perReq x reqs
- Display
- 1 or more: up to 2 decimals with thousands separators; under 1: exactly 4 decimals
What it assumes
- Every request is the same size: one average for input tokens and one for output tokens.
- The three presets are generic examples, not quotes from any vendor: Economy ($0.8 in, $4 out per million tokens), Balanced ($3, $15) and Frontier ($15, $75). The tool starts on Balanced with 2,000 input tokens, 500 output tokens and 10,000 requests.
- Blank, negative or non-numeric entries count as 0.
- Prompt caching, batch discounts, reasoning or tool-call tokens, retries and infrastructure are not modeled, and the tool does not count tokens from text.
- There is no notion of a month. For a monthly figure, enter the requests you expect in a month.
Worked example
Illustrative only: an internal assistant with a prompt of about 1,500 tokens (instructions plus retrieved text) and a reply of about 400 tokens, at 20,000 requests in a month. The Balanced preset prices are a starting point, not a quote from any provider.
| Input | Value |
|---|---|
| Input price | 3 $ / 1M tok |
| Output price | 15 $ / 1M tok |
| Input tokens / request | 1500 |
| Output tokens / request | 400 |
| Number of requests | 20000 |
- Input cost per request: 1,500 / 1,000,000 x 3 = 0.0045
- Output cost per request: 400 / 1,000,000 x 15 = 0.0060
- Cost per request: 0.0045 + 0.0060 = 0.0105, shown as $0.0105 (under $1, so four decimals)
- Total: 0.0105 x 20,000 = 210, shown as $210
Cost per request $0.0105; total for 20,000 requests $210. Output is 21% of the tokens but 57% of the cost, because its rate is five times the input rate.
Frequently asked questions
- Are the preset prices current?
- No. They are generic example tiers and provider prices change often. Check your provider's pricing page and type the real rates into the two price fields.
- How do I get the token counts?
- Take the usage figures your provider returns with each response, or use its token counter, on a sample of real traffic and average them. Different models can count the same text differently.
- How close will this be to my real bill?
- It is an order-of-magnitude budget, not a forecast. Spread in request sizes, retries, caching, batch discounts and reasoning tokens all move the real figure. Compare the estimate with a week of actual usage.
- Does it account for prompt caching?
- No. It applies one input price to all input tokens. If part of your prompt is cached at a different rate, estimate the cached and uncached portions separately and add the results.
- Why does output often dominate the cost?
- Providers commonly charge more per output token than per input token. Change the price and token fields one at a time to see which one drives your total.
Read the guide
Building this for real?
I help teams design and ship AI systems — cost modelling, RAG, agents, and architecture.
Work with me