LLM cost per 1,000 requests: a plain-English breakdown
There's no fixed "price per request" for an LLM API — you pay per token, so cost per 1,000 requests depends entirely on how many tokens each request uses. Here's the one-line formula, a table of what 1,000 requests actually costs across models and usage sizes, and the mistake that makes people overestimate (or underestimate) their bill.
The formula
cost per 1,000 requests = 1,000 × [ (avg input ÷ 1,000,000 × input price) + (avg output ÷ 1,000,000 × output price) ]In words: work out the per-request cost from your average input/output tokens and the model's per-token rates, then multiply by 1,000. That's it. The only inputs you need are your average tokens per request (in and out) and the model's price.
Example: a request with 1,000 input + 500 output tokens on a $1 / $5 (per 1M) budget-Claude-class model costs (1000÷1e6×1) + (500÷1e6×5) = $0.001 + $0.0025 = $0.0035. Times 1,000 = $3.50 per 1,000 requests.
Cost per 1,000 requests, by model & size
Three common request shapes — short (chat), medium (RAG), large (document) — across tiers. Figures are per 1,000 requests.
| Model | Short 500 in / 150 out |
Medium 2k in / 400 out |
Large 6k in / 800 out |
|---|---|---|---|
| Budget (mini/flash class) | ~$0.13 | ~$0.54 | ~$1.38 |
| Claude Haiku 4.5 | $1.25 | $4.00 | $10.00 |
| Claude Sonnet 5 | $3.75 | $12.00 | $30.00 |
| Claude Opus 4.8 | $6.25 | $20.00 | $50.00 |
Do it for your exact usage
Enter your average tokens and see cost per request, per 1,000, and per month across every model — sorted cheapest first.
Open the calculator →The mistake that skews the number
Counting requests instead of tokens
The most common estimating error is treating all requests as equal. As the table shows, a "large" request can cost 8× more than a "short" one on the same model — because it carries 12× the tokens. If your traffic is a mix, an average based on request count alone will be wrong. Use your average tokens per request, and if the mix is very uneven, estimate the buckets separately and add them.
Forgetting output is the expensive side
Output tokens cost ~4–5× input. A short prompt that triggers a long answer can cost more than a long prompt with a one-line answer. Watch output length, and cap max_tokens where you can.
Lowering your cost per request
Three fast levers (full list in how to cut your bill 50%+):
- Route the request to a cheaper model if the task allows — the biggest single change.
- Trim input — leaner retrieval and bounded chat history cut the input side.
- Cache the repeated prefix so you stop re-paying for a fixed system prompt.
FAQ
How do I calculate LLM cost per 1,000 requests?
1,000 × [(avg input ÷ 1M × input price) + (avg output ÷ 1M × output price)]. Per-request cost from your token averages, times 1,000.
Why is my cost per request higher than expected?
Usually because request count hides token variation — a context-heavy or long-output request costs many times more than a short one. Estimate from average tokens, not request count.