LLM API Cost Calculator
Price a feature before you ship it. Token counts and request volume in, monthly bill out, with prompt caching and batch discounts modelled properly.
Model and prices
Prices are per million tokens and every field is editable. Paste in whatever your provider actually charges you.
Your workload
One request means one round trip to the model. A multi-turn conversation resends its history, so count the cumulative input.
About 3K words.
About 450 words.
The share of your input that is a stable, repeated prefix.
Around 10% on most providers.
50% is standard for asynchronous batch endpoints.
Failed calls you pay for and re-run. 5% to 15% is typical in production.
Monthly cost
$13,230
$0.02 per request · 92M tokens a day · $160,965 a year
Per request
$0.02
Per day
$441
Output share of bill
43%
Input is the other 57%.
Cost per 1k requests
$22
The number to quote in a design doc.
Input dominates at 57% of the bill, which is the profile that prompt caching was built for. You have caching set to zero. If any part of your prompt repeats across requests, raise the hit rate slider and watch what happens.
The same workload on every model
Monthly cost at your volume, holding token counts and discounts constant.
Your current selection.
Routing this workload to Open-weights (hosted) (Various) instead would cost $1,096 a month, a saving of $12,134. That is only the right call if the smaller model actually clears your quality bar, which is a question for an eval set and not for a pricing page. The usual answer in production is to route by difficulty and send only the hard minority to the expensive model.
Scaling out
| Requests per day | Per day | Per month | Per year |
|---|---|---|---|
| 1,000 | $22 | $662 | $8,048 |
| 10,000 | $221 | $6,615 | $80,483 |
| 100,000 | $2,205 | $66,150 | $804,825 |
| 1,000,000 | $22,050 | $661,500 | $8,048,250 |
| 10,000,000 | $220,500 | $6,615,000 | $80,482,500 |
Default prices are published list rates as of the date on this page and are editable above. They exclude volume commitments, regional pricing, fine-tuning and any enterprise agreement, and they change often. Check your provider’s pricing page before putting a number in a budget.
Key takeaways
- LLM API pricing is quoted per million tokens and charged separately for input and output, with output typically costing three to five times what input costs on the same model.
- Because output is the expensive side, capping response length is usually a bigger lever on the bill than shortening the prompt.
- Prompt caching charges repeated input tokens at a large discount, commonly around a tenth of the base input rate, which makes a long fixed system prompt across many requests far cheaper than the raw token count suggests.
- Batch APIs typically apply a 50% discount in exchange for asynchronous delivery, so any workload that does not need an immediate response should not be paying interactive rates.
- A token is roughly 0.75 English words, so 1,000 tokens is about 750 words and a 100,000-token context is about 75,000 words.
Frequently asked questions
- How do you calculate the cost of an LLM API call?
- Multiply input tokens by the input price per million and divide by a million, do the same for output tokens at the output price, and add the two. A request with 2,000 input tokens and 500 output tokens on a model priced at $3 per million input and $15 per million output costs (2000 ÷ 1,000,000 × $3) + (500 ÷ 1,000,000 × $15) = $0.0135. Multiply by request volume for the daily or monthly figure.
- How many tokens is a word?
- Roughly 1.3 tokens per English word, or about 0.75 words per token. A useful rule of thumb is that 1,000 tokens is about 750 words, or a page and a half of prose. Code, JSON and non-English text tokenise less efficiently, so budget 20% to 50% more for those.
- Does prompt caching actually reduce costs?
- Substantially, when the repeated part of your prompt is large and stable. Cached input tokens are typically billed at around a tenth of the normal input rate, with a one-time write premium on the first call. The break-even usually arrives after two or three reuses, so it pays off for a long system prompt or a fixed document that many requests share, and does not pay off for prompts that change every call.
- What is the cheapest way to run a high-volume LLM feature?
- Route by difficulty rather than defaulting everything to the largest model. Most production traffic is classification, extraction and short summarisation, which a small model handles at a fraction of the cost, leaving the frontier model for the minority of requests that actually need it. Combine that with prompt caching for shared context and a batch endpoint for anything asynchronous.
- Why is my actual bill higher than the calculator says?
- The usual causes are retries after failures, multi-turn conversations resending the entire history as input on every turn, agent loops that make several model calls per user action, and reasoning tokens that are billed as output but never shown to the user. Model a conversation as its cumulative input, not as one request, and add a retry factor of 5% to 15% for anything user-facing.
- Are these prices current?
- The defaults reflect published list prices at the date shown on this page, and every price field is editable so you can paste in whatever your provider currently charges. Model pricing changes often and volume commitments, regional rates and enterprise agreements all move the number, so treat the defaults as a starting point rather than a quote.