Kimi K3.1 Pro API pricing and cost calculator

How this is calculated

Start with the calculator's default request. With 100,000 input tokens and 5,000 output tokens, uncached input costs $0.20 and output costs $0.05, so one call comes to $0.25. Read half the input from cache and the same call costs $0.16. At 100 calls a month that's $16. Real jobs rarely look like that, so here are three that do. An agent session is the first. Picture 40 turns where every turn resends a 150,000 token prompt and writes 3,000 tokens back, with most of the prompt read from cache. That session costs $3.48. If the cache never hits, the same session costs $13.20. Cache reads are 90% cheaper than uncached input, and an agent loop is where that adds up, because the prefix is the same on every turn. The second job is a nightly document pipeline. Say 2,000 documents each night, with 30,000 input tokens and 1,500 output tokens per document and nothing cached. One night costs $150, and 30 nights of it cost $4,500. Nothing is cached because every document is new, so the input rate is the number that matters. The third is a reasoning heavy task with a short prompt and a long answer, such as 20,000 input tokens and 30,000 output tokens. That call costs $0.34. Output is billed at $10 per million tokens while input is $2, so on jobs like this the answer length deserves as much attention as the prompt, and a longer reasoning budget shows up on the bill. Image and video inputs are billed by the provider's own token counting, so run a sample file through the API and read the usage it reports before you budget a media pipeline.

Verdict

Kimi K3.1 Pro is worth a test when you want a large open weight model for agent work and long documents, and when the weights being open matters to you. Run it on your own tasks and record accepted results, not just token counts. Count the turns an agent needs, how much of each prompt hits the cache and how long the reasoning output runs. Read the model license before any commercial use, since it's a custom license and this page doesn't summarize it. A third party host also lists a lower price right now, so price the host you would really call and check its weight precision, output limit and uptime.