LLM API prices, explained: per-million tokens, context windows and free tiers

API prices for language models are quoted per million tokens, and every model has two of them: one for the tokens you send (input, or prompt) and one for the tokens the model writes back (output, or completion). Output is almost always more expensive, often by three to five times, because generating a token costs more compute than reading one. A price of $0.25 in and $1.00 out means a request with 3,000 tokens of context and 500 tokens of answer costs about $0.0013.

We collect these prices every day for every model on OpenRouter, a router that exposes hundreds of models from many providers behind one API. The table on the model prices page is that snapshot, and the daily API prices report is written from the difference between one day and the next.

A token is a chunk of text, usually a word or part of a word. In English, a thousand tokens is roughly 750 words; code and other languages tokenize less efficiently, so the same text can cost more. Prices per million tokens are comparable across models only if you remember that different tokenizers split the same text into different counts. The site does not convert between tokenizers; the prices are as the provider quotes them.

The context window is the maximum number of tokens a model can hold at once, prompt and answer together. A 1,000,000-token window is not free to use: every token in the window is billed as input on every request, so long conversations get expensive linearly. The context column exists so you can tell whether a model can take your documents at all, not as a measure of quality.

Some models on OpenRouter are listed at $0 for both input and output. We flag them as free only when both prices are exactly zero; a model with no published price is unknown, not free, and is shown that way. Free tiers come with conditions the price does not show: rate limits per minute and per day, lower priority when the provider is busy, and in some cases the provider's right to use your prompts for training. The free models page tracks which models are free on a given day and the changes page shows what entered and left the free tier, because those lists change weekly.

A model marked as open weights has its weights published on Hugging Face, which means you can run it yourself instead of paying the API. That is the only case where the API-or-GPU comparison applies. Closed models have one price, the API's, and no alternative to compare it with.