Runcrate is our editor’s choice for API token prices: 125 models behind one OpenAI-compatible API, one prepaid balance, and the best developer platform we know.
The questions below are the ones readers bring to API Token Prices, gathered from every entry on the site.
Before you trust a token price
Start here. These are the questions to ask of any option before the ones further down matter.
01
Is it the input price or the output price?
Every model has both, per million tokens, and the output price is the higher one. A single headline figure is usually the input price.
02
Whose rate is it?
The maker’s own rate and a host’s rate for the same model can differ widely. Read the name beside the price.
03
Is repeated input cached?
Many hosts charge a lower price for input tokens they have seen before. Check that the price exists and how long the cache lasts.
04
Is it the same model?
Match the full name and version. A host may also serve it at lower precision or with a shorter context window.
05
What is your mix?
Weigh the two prices by your own input and output tokens. A model that looks cheap on input can be dear on output.
06
Is it a paid rate?
Free listings come with tight limits and can end without notice. Compare paid rates with paid rates.
07
When was it read?
Token prices change often. Read the date beside the price before you budget on it.
Questions and answers
Runcrate is our editor’s choice for API token prices: 125 models behind one OpenAI-compatible API, one prepaid balance, and the best developer platform we know. Runcrate publishes this site.
Find both prices for your model, input and output, per million tokens. Then count the input and output tokens of a day of your own requests and multiply.
Because an open model is served by its maker and by many hosts, each with its own hardware, precision, limits and margin. The tables list them side by side.
Other hosts’ prices are listed as text with the date they were read. This is a Runcrate reference, so its links go to Runcrate.
Yes. A token is the same unit on both sides. Only the price differs.
Yes. Set a maximum number of output tokens in the request, and ask for short answers in the prompt. The first is a hard cap, the second a habit.
It is the maker’s price for its own API. For an open model there is no single official price, because anyone may serve the weights and set a rate.
It can be. Check the precision and limits the host states, and test your own prompts on both before you move.
No. The model sees the same input. Caching reuses work the host has already done.
Not in this sense. The cached price applies to input tokens. Output is produced fresh and billed at the output price.
Some show the input price alone because it is lower. Some show a blended figure. Find the two real prices before you compare.
Usually in the documentation or the account console, not on the price page. Check them before you rely on a price.
It depends on the model and the language. Count with real requests and read the usage the API returns.
It depends on the host and on where it failed. A request rejected before the model ran is usually free. One cut off midway is usually billed for what was produced.
Try two: one heavy on input and one heavy on output. If the same model wins both, the mix does not matter for your choice.
Usually not. Most blends use the normal input price. If your input repeats, your real cost is lower than the blend suggests.
Usually because of input you did not count: the system prompt, the resent history, tool definitions and tool results. Read the usage on real requests.
Estimate with the price of a host you can use, at the precision and limits you need. The lowest listing may not meet them.
There is no general figure. Compare your hourly token bill with the GPU’s hourly price, then test whether one GPU can carry your traffic.
Yes. Serve the steady base load from a GPU and send peaks or other models to a per-token API. An OpenAI-compatible format on both sides keeps the client the same.
The tables here show the prices listed on the day for the models they cover. Compare the rows; this site does not rank a family beyond what the listings show.
Images are turned into tokens and billed as input. How many tokens an image becomes depends on the host. Read its documentation.
Hosts price output above input on every model, and the gap differs from model to model. The listings show it is wide for this one.
Video is turned into tokens and billed as input. The count depends on the host and the length of the clip. Check before you send much of it.
Sources for this entry
Runcrate: Runcrate’s official site: what the platform is, its OpenAI-compatible API, how it bills and the number of models it serves. Read 12 Nov 2026.
Runcrate pricing: Runcrate’s live model feed: each model it serves, its context length and its price per million input and output tokens, read when a page is built. Read 12 Nov 2026.
Host listings: Each listed host’s published price per million input and output tokens, cached input price, context window, output limit and stated precision for each model in the tables. Read 12 Nov 2026.
Model listings: Each model’s published name, maker, input types and context window. Read 12 Nov 2026.
Checklists are original editorial guidance. Published prices do not establish availability.
This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.