LLM API Cost Calculator
Estimate your monthly bill across GPT, Claude, Gemini, Grok, Mistral, DeepSeek, Qwen, Kimi, GLM, Doubao, Hunyuan, MiniMax and Xiaomi MiMo from requests per day, average input and output tokens, and how much of your prompt is cached. Sorted cheapest first.
TL;DR: Enter requests per day, average input/output tokens and cache-hit share to see the daily and monthly API bill for every listed model, cheapest first, using official list prices.
Start from a typical workload:
Estimated cost, cheapest first
| Model | $ / 1M in | $ / 1M cached | $ / 1M out | Per request | Per day | Per month | Source |
|---|
Prices checked 2026-10-22 on each vendor’s official page (USD, standard tier). The tier for long prompts is chosen from your average input tokens. Cache-write fees, batch discounts, taxes and free tiers are not included.
How to use the cost calculator
- Measure a few real requests: paste a typical prompt (system prompt, tools, history included) into the token counter to get average input tokens, and read output tokens from the API’s usage field.
- Enter requests per day and, if you use prompt caching, the share of input you expect to hit the cache.
- Untick models you would never use, then compare the monthly column. Re-check the vendor page before you commit a budget.
Related guides: Estimate LLM API costs (worked examples) · API price comparison · Prompt caching basics · MiMo V2.5 → V2.6 migration · How to get LLM API keys
FAQ
How is the monthly cost calculated?
For each model: (uncached input tokens × input price + cached input tokens × cached-input price + output tokens × output price) ÷ 1,000,000 gives the cost per request. That is multiplied by requests per day and by the number of days you enter (30 by default).
Where do the prices come from?
The same price table as our token counter, copied by hand from each vendor’s official pricing page (checked 2026-10-22). Each row links to its source. Promotional prices and scheduled price changes are noted next to the model.
What does “cached input %” mean?
Most vendors bill a repeated prompt prefix (system prompt, tool definitions, long documents) at a lower “cache hit” rate. Enter the share of your input tokens you expect to be served from cache. Cache-write surcharges are not included, so treat the result as a steady-state estimate.
Are thinking / reasoning tokens included?
Only if you include them in “average output tokens”. Vendors generally bill reasoning tokens at the output rate, so add them there.
Why do token counts differ between vendors?
Each vendor has its own tokenizer. If your averages were measured with an OpenAI tokenizer, tick the Claude adjustment: Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text.