LLM API Cost Calculator
Enter your usage once — see what it costs per month on every major model. No sign-up, runs in your browser.
| Model | Provider | Input $/1M | Output $/1M | Per request | Per day | Per month |
|---|
How this calculator works
Every LLM API bills the same way: a price per million input tokens (your prompt, system message, conversation history, retrieved documents) and a separate, higher price per million output tokens (the model's response). Your monthly bill is simply:
(input tokens × input price + output tokens × output price) × requests
The hard part is that every provider publishes prices on a different page, in a different format, with different caching rules — so comparing them honestly takes an afternoon. This table does that comparison in one place and we keep the numbers current as providers change them (which they do almost monthly).
What usually surprises people
- Output tokens dominate. Output is typically 3–5× the input price. A chatbot that writes long answers costs far more than one that writes short ones, even with identical traffic.
- Prompt caching is the biggest lever. If every request shares a long system prompt or document, cached-input pricing (often 90% off) can cut the input side of your bill dramatically. Use the "cached input share" selector above to model it.
- The spread is enormous. Between a frontier flagship and a budget model there is often a 50–100× price difference. Many production workloads run fine on mid-tier models.
FAQ
How do I estimate tokens per request?
Rule of thumb for English: 1 token ≈ 4 characters ≈ 0.75 words. A typical chat message is 50–200 tokens; a page of text is ~500 tokens; a long system prompt with instructions can be 1,000–3,000 tokens. Most providers ship a tokenizer or a count_tokens endpoint if you need exact numbers.
Are these prices up to date?
Yes — this table syncs automatically on a schedule from a live pricing feed that tracks each provider's current per-token rate, including active promotions. That's the point: a hand-maintained comparison goes stale within weeks, this one doesn't. The "last synced" date above is the last automated run, not a one-time snapshot. Regional endpoints and batch APIs can still be cheaper — footnotes under the table flag the big ones.
Why is my real bill higher than this estimate?
The usual culprits: conversation history re-sent on every turn (input grows with each message), reasoning/thinking tokens billed as output on some models, retries, and tool-calling round trips. Multiply your naive estimate by 1.5–3× for a multi-turn assistant.
Which model should I pick?
Price this table can tell you; quality it can't. A sane default: prototype on a flagship model, then step down tier by tier until quality degrades, and stay one tier above that.