How to estimate LLM API costs
- Enter the average number of input tokens (prompt, system message, documents and history) and output tokens (the reply) per request. Use our Token Counter on a typical prompt to get realistic numbers.
- Enter how many requests you expect per day. Add the share of input that is served from a prompt cache, and whether you can use a batch API.
- The table ranks models from cheapest to most expensive per request, day and month. Edit any price, delete rows, or add your own model; your price list is remembered in this browser.
What drives the cost
LLM APIs charge per million tokens, with output tokens typically four to five times more expensive than input. That means a chatbot with long answers costs more than a classifier with one-word replies, even with the same prompt. Long system prompts and retrieved documents are repeated on every call, which is where prompt caching helps: providers charge much less for input they have recently seen, often 10% to 50% of the normal price. Batch APIs, for jobs that can wait up to a day, usually cut prices by half.
The default prices are published list prices for standard context lengths, last checked on 20 October 2026 for every provider: OpenAI (GPT-6 and GPT-5 families), Anthropic (Claude 5.5), Google (Gemini 4 Argon, Gemini 3.8 Flash), Mistral and DeepSeek. Where a provider has announced a price change, the table switches to the new price on that date: Gemini 4 Argon doubles on 1 December 2026 and Gemini 3.8 Flash on 1 January 2027. DeepSeek prices are peak-hour rates; its off-peak rates are half. Mistral does not list per-token prices for its newest models on its pricing page, so those rows use the launch list prices. Long-context surcharges, reasoning tokens and launch discounts are not included, so check the provider's pricing page before committing a budget.
Why are output tokens more expensive?
Output is generated one token at a time, which uses far more computation per token than reading the input in parallel.
How can I reduce my LLM costs?
Shorten system prompts, cache repeated context, cap output length, route simple requests to smaller models, and use batch processing for non-urgent work.
Are the prices up to date?
All default prices were checked on 20 October 2026, and announced price changes switch over automatically on their date. Prices change often, so edit any row to match the provider's current pricing page, or add the models you use. Your edits are saved in this browser.