Tokens and usage
Token counts, API usage fields, and the free preview.
The model reads and writes tokens, pieces of words. This preview's model price is zero. Token usage still counts toward the rate limits and the 262,144-token context window.
How many tokens
Roughly, in raon-1-mini-preview:
| Text | About |
|---|---|
| Korean | 1 token per 2 to 3 characters |
| English | 1 token per word |
Treat these as estimates for planning. The usage field of each response gives the exact count.
Usage in responses
| API | Input | Output | Of which reasoning |
|---|---|---|---|
| Chat Completions | usage.prompt_tokens | usage.completion_tokens | usage.completion_tokens_details.reasoning_tokens |
| Responses | usage.input_tokens | usage.output_tokens | usage.output_tokens_details.reasoning_tokens |
| Messages | usage.input_tokens | usage.output_tokens | not reported separately |
The model's reasoning counts as output tokens and uses the same output budget as the answer (max_tokens, max_completion_tokens or max_output_tokens, depending on the API). Reasoning is always enabled; leave room for it when setting the output limit.
Messages reports cache reads and writes separately in cache_read_input_tokens and cache_creation_input_tokens. Its input_tokens excludes those cached tokens. Chat Completions and Responses include cached input in their total input count and report cache details separately when available.
When streaming, usage comes at the end: in a final chunk on Chat Completions when you set stream_options.include_usage, in the response.completed event on Responses, and in message_start and message_delta on Messages.
If a stream is interrupted, final usage may be missing. Raon uses the model server's usage when available; otherwise it estimates usage from the input and output text it received. These estimates can differ from the model's token counts. Web search can run several model turns; usage includes all of them.
Price
The model is free during this preview. GET /v1/models exposes pricing fields in US dollars per million tokens:
"pricing": {
"input_per_mtok": "…",
"output_per_mtok": "…",
"cache_read_per_mtok": "…",
"cache_write_per_mtok": "…"
}For priced models, ordinary input, cache reads, cache writes and output each use their applicable rate. The input-plus-output formula applies when there are no cached tokens. Requests refused with an error status are not billed; streams that start successfully can still record usage if interrupted. During this preview, the model price remains zero in either case.