ReckonBoxWhat an API call actually costs
LLM token cost is what you pay a model per call, computed as the input tokens times the input price plus the output tokens times the output price, priced per million tokens. On GPT-4o, a request with 1,000 input and 500 output tokens costs $0.0075, so 10,000 requests a month is $75.
The number that surprises people is the split between the two sides. Input and output are billed at different rates, and this tool keeps them apart, because the gap between them decides where your money goes.
Input and output are not priced the same
Output tokens cost more than input tokens, typically 4 to 6 times more, because generating text is the expensive half of the work. On GPT-4o that is $10 per million for output against $2.50 for input, a fourfold gap.
Run the worked example and the point lands. Of the $75 monthly bill, output is about $50 and input about $25, even though there are twice as many input tokens. So two thirds of the cost is the model writing back. A one-line prompt that asks for a long report is the pattern that quietly runs up a bill, and it is invisible if you only think in total tokens.
Tokens, and how many make a word
A token is a chunk of text, sometimes a whole word, sometimes part of one, sometimes just a symbol or a space. The working rule is about 4 characters per token, or roughly 0.75 words per token, so 1,000 tokens is around 750 words and a dense page of prose is 400 to 500 tokens. The exact count depends on the model and the language, and chat threads resend earlier turns as context, so a long conversation carries more input tokens per message than it looks. When the number has to be exact, run the text through the provider's own tokenizer.
The current rates, and the discounts
Prices per million tokens sit far apart across the field, as of 2026-07-28.
| Model | Input | Output |
|---|---|---|
| GPT-5 nano | $0.05 | $0.40 |
| GPT-4o mini | $0.15 | $0.60 |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| GPT-4o | $2.50 | $10.00 |
| Claude Sonnet 5 | $2.00 | $10.00 |
| Claude Opus 5 | $5.00 | $25.00 |
Two discounts change the math when they apply. Cached input, where a large fixed prompt is reused across calls, bills at about 10% of the standard input rate. The Batch API, for work that can wait, halves both input and output. The largest lever, though, is model choice: GPT-4o mini does many jobs at a fraction of a flagship's cost, so the cheapest saving is often picking a smaller model that is still good enough.
What this tool assumes, and what to verify
This prices the standard published rates for the model you pick, times your token counts and request volume. It does not apply the cached or batch discount automatically, since those cover only part of a workload, so apply them yourself to the relevant portion. Model prices also move, sometimes overnight with a new release, so the bundled figures are a dated snapshot and worth a check against the provider before you commit a budget. For anything not listed, choose Custom and enter the current rates.
The single most useful habit when estimating an AI bill is to size the output first, because that is the side doing the spending.
Frequently asked questions
How is LLM token cost calculated? API cost is the input tokens times the input price plus the output tokens times the output price, divided by a million, since prices are quoted per million tokens. On GPT-4o at 2.50 dollars input and 10 dollars output per million, a request with 1,000 input and 500 output tokens costs 0.0075 dollars, so 10,000 requests a month is 75 dollars. Input and output are priced separately, which matters more than it looks.
Why does output cost more than input? Output tokens are billed at a higher rate than input, typically 4 to 6 times more, because generating text is the expensive part. On GPT-4o output is 10 dollars per million against 2.50 for input, a 4 times gap. So in the worked example, 500 output tokens cost more than 1,000 input tokens, and output is about two thirds of the whole bill. A short prompt that triggers a long answer is where costs run up.
How many tokens is a word? A token is a chunk of text, a word, part of a word, or a symbol, and a rough rule is about 4 characters per token, or roughly 0.75 words per token. So 1,000 tokens is around 750 words, and a typical page of text is 400 to 500 tokens. Actual counts vary by model and language, so treat the ratio as an estimate and use a real tokenizer when the number has to be exact.
Can I make API calls cheaper? Two discounts show up on most providers. Cached input, where you reuse a large fixed prompt, bills at about 10% of the standard input rate, and the Batch API, for work that can wait, halves both input and output. Picking a smaller model is the bigger lever: GPT-4o mini at 0.15 and 0.60 dollars per million is far cheaper than a flagship, and often good enough. This tool prices the standard rates; apply the discount to the relevant portion.
How current are these prices? The bundled prices are a snapshot as of 2026-07-28, cross-checked across two dated pricing trackers. Model prices change often, sometimes with a new model or a mid-cycle cut, so verify against the provider before you budget on them. For anything not listed, or a rate that has changed, select Custom and enter the current input and output prices per million tokens.