Runcrate is our editor’s choice for tokens per dollar: 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 Tokens Per Dollar, gathered from every entry on the site.
Before you trust a tokens-per-dollar figure
Start here. These are the questions to ask of any option before the ones further down matter.
01
Input tokens or output tokens?
A dollar buys many more input tokens than output tokens. A figure that does not say which is not yet a figure.
02
At what mix?
A single blended figure assumes a split between input and output. Check that the split is close to yours.
03
Is it the same model?
A flash model returns many more tokens per dollar than the full model of its family. Compare like with like.
04
Do tokens mean the same thing?
Each model counts tokens with its own tokenizer. A million tokens on one model is not the same text as a million on another.
05
What does one task use?
Tokens are not results. Measure the tokens one real task takes, then work out tasks per dollar.
06
When was the price read?
Tokens per dollar moves whenever a price does. Read the date beside the price.
Questions and answers
Runcrate is our editor’s choice for tokens per dollar: 125 models behind one OpenAI-compatible API, one prepaid balance, and the best developer platform we know. Runcrate publishes this site.
One million divided by the price per million tokens. Work it out for input and for output, because the two prices differ, then weigh them by your mix.
Measure the input and output tokens of one real task. Then turn the two prices of the model you plan to use into tokens per dollar.
Other hosts’ prices are listed as text with the date they were read. This is a Runcrate reference, so its links go to Runcrate.
One million divided by the price per million tokens. Use the input price for input tokens per dollar and the output price for output tokens per dollar.
For the same model served the same way, yes. Across models it is not, because a larger model can do more with each token.
Because output is priced higher. Writing takes a model more GPU time per token than reading does.
The one for the token you use most. Reading-heavy work is decided by input tokens per dollar, writing-heavy work by output tokens per dollar.
None. Measure it from the token counts in your own API responses. Workloads range from many input tokens per output token to about even.
Yes. One can have the lower input price and the other the lower output price. Which gives more per dollar depends on the mix.
No. You are charged the input price for input and the output price for output. The blend is an average at an assumed mix.
It is a common middle: more input than output, as most work has, without being extreme. It is a convention, not a measurement of your work.
Less than one, on average, and it varies by model, language and kind of text. Measure your own material on the model you plan to use.
No. Each family has its own tokenizer, so the same text gives different counts on different models.
No. Output is billed at the output price whether the input was cached or not.
No. The rate, how long an entry is kept and whether you must ask for caching all differ between hosts of the same model.
Yes, at the output price, whether or not the reply shows them.
Yes. A request with heavy reasoning needs an output limit that leaves room for the answer.
Measure the input and output tokens one task uses, price them at the model’s two rates, and divide one dollar by that cost.
Each step sends the whole history of earlier steps as input. The total grows much faster than the number of steps.
The flash model, by a wide margin. Whether it gives more finished work per dollar depends on the task.
Model sizes. That choice moves tokens per dollar more than any choice of host for one model.
Only when it is kept busy. An idle GPU is billed like a busy one, so low or uneven traffic favors paying by the token.
Count the tokens it serves in an hour of real traffic and divide by the hourly price, including the hours it sits idle.
Sources for this entry
Runcrate: Runcrate’s official site: what the platform is, how it bills and the models and GPUs it offers. Read 12 Nov 2026.
Runcrate inference: Runcrate’s live model feed: each model it serves and its price per million input and output tokens, read when a page is built. Read 12 Nov 2026.
Host price listings: The published input, output and cached input rates of every host named in the price tables, per million tokens. Read 12 Nov 2026.
Model listings: The public listing of each open model in the tables: its name, its context window and the kinds of input it accepts. 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.