AI API Cost Calculator
Describe your workload once — get the monthly cost for any model, with caching and batch discounts included. Prices updated 2026-08-02.
- Per request
- $—
- Per day
- $—
- Input share
- Savings from discounts
- $—
Same workload, every model
Cheapest first. Recalculates as you edit the inputs above.
| Model | Per request | Per month | vs cheapest |
|---|
Worked example: a support chatbot at 1,000 requests/day
Say each request carries a 1,200-token system prompt + history and 300 tokens of user input (1,500 in), and the bot answers with 400 tokens. On a $5/$25-per-million flagship model:
- Input: 1,500 ÷ 1M × $5 = $0.0075 · Output: 400 ÷ 1M × $25 = $0.0100 → $0.0175/request
- With caching (70% of input cached at −90%): input drops to ~$0.0028 → $0.0128/request
- 1,000 requests × 30 days = ~$384/month (vs $525 uncached — caching saved 27%)
- The same workload on a "mini" tier model at $0.25/$1.00: ~$19/month
That last line is the most important pricing lesson in AI: model choice moves your bill 20×, caching moves it ~2×, everything else is noise.
Frequently asked questions
› How do I estimate tokens from words or requests?
A useful rule of thumb: 1,000 tokens ≈ 750 English words. A typical API request with a system prompt and a short user message runs 500–2,000 input tokens, and a paragraph-length answer is 150–400 output tokens. When in doubt, measure a few real requests in your provider dashboard and use the averages here.
› Why is output so much more expensive than input?
Generating tokens requires a full forward pass of the model per token, while input tokens are processed in parallel. Providers price that compute difference in: output typically costs 3–5× the input rate. Workloads that generate long answers (writing, code generation) are output-heavy and cost more than classification or extraction tasks.
› How much does prompt caching actually save?
Cached input tokens typically cost 90% less than fresh ones. Chatbots and agents resend the same system prompt and conversation history with every request, so 60–90% of their input volume can be cache hits. For a busy chatbot this single feature often halves the total bill — enable the caching toggle in the calculator to see the effect on your numbers.
› Are these prices accurate?
Prices are refreshed automatically from public provider listings (the date is shown above the results). Providers occasionally change prices with little notice, and enterprise contracts get custom rates — always confirm on the provider’s official pricing page before locking a budget.
› Which model is cheapest for my use case?
Use the comparison table below the calculator: it recalculates your exact workload across every model we track and sorts by monthly cost. As a rule, "mini/flash/haiku" tier models cost 10–30× less than flagship models and handle routine tasks (classification, extraction, simple chat) well. Reserve flagship models for the hard 10% of requests.