What is token-based pricing?
Token-based pricing means an API charges for each token it processes, with one rate for the tokens you send in and another, usually higher, rate for the tokens the model writes back. The cost of a request is input tokens times the input rate plus output tokens times the output rate. You pay for how much text moves through the model, not for how long a machine runs.
Input tokens and output tokens
A token is a chunk of text, roughly three-quarters of a word in English, that a model processes as one unit. Input tokens are everything sent to the model: the prompt, the system instructions and, in a multi-turn conversation, every earlier turn resent with each new request. Output tokens are what the model generates.
Most vendors price the two separately, and output tokens often cost three to five times as much as input tokens. The reason is mechanical: the model reads the whole input in one parallel pass, but it generates output one token at a time, each a separate sequential step.
How cost scales with context length
A multi-turn conversation usually resends the full transcript as input on every turn, so the input count grows with each exchange, not only with the length of the newest message. By turn ten you have paid for the first message ten times. Long documents and large retrieved-context payloads make this worse: paste a 20-page contract into the prompt and you pay the input rate on every token of it, every time it is sent.
Why idle time costs nothing on a token meter
Token pricing charges only for tokens processed, so an API that receives no requests costs nothing in that hour. For spiky or low-volume workloads that is a real advantage. The catch is the rate itself: the vendor sets it, it includes the vendor's margin and infrastructure cost, and it does not fall as your volume grows unless the vendor offers a separate volume tier.
An alternative: flat hourly compute
Renting a GPU by the hour pays for the same work a different way. You pay a fixed rate for the time a machine is assigned to you, and you can push as many tokens through it as it can handle in that window. Cost becomes predictable and independent of volume. The flip side is that idle time between requests is time you still pay for, which a token meter avoids.
Which model costs less depends on how steady and how heavy your usage is. A useful test: divide the hourly rate by the tokens per hour you would really push through the machine, and compare that number with the API's per-token price for the same model class.
Where GPUwerk fits
GPUwerk rents dedicated DGX Spark machines by the hour, billed per minute, not per token: $0.79/hour for one Spark, or $1.79/hour for a linked pair with 256 GB of combined memory. The pricing page has the numbers for both. For what a token costs to process at the hardware level, see the glossary entries on prefill and decode.