Kimi K3.1 Pro: lower token rates, similar measured task cost

In its October 23 evaluation, Artificial Analysis reports about $1.98 per task for K3.1 Pro, roughly matching K3 despite token rates falling by 33%. It records 55,000 output tokens per Intelligence Index task, versus 48,000 for K3, a 14% increase. These figures describe that evaluation workload, not a guaranteed customer bill.

For developers estimating usage, Kimi’s Chinese-language billing documentation defines tokens as character sequences; one MTok means one million tokens. Both supplied input and generated output incur usage charges. For K3 models, writing reusable input into the cache is charged separately according to its retention period, either five minutes or one hour. A later cache hit does not incur another write charge.

The cache guide recommends placing unchanged instructions and reference material before changing questions. Reuse matches the request’s beginning: a change prevents reuse beyond that position.

For Chat Completions accounting, the same guide identifies usage.prompt_tokens as total input, with cached_tokens and cache_write_tokens inside prompt_tokens_details separating cache reads and writes. Streaming requests need stream_options.include_usage=true for complete usage details in the final chunk.

THE DECODER also examines the release’s pricing and task-cost comparison in its coverage.