Kimi K3.1 Pro: API costs and Highspeed choices

The direct international API tariff below is in USD per million tokens, checked October 26, 2026, excluding applicable taxes. Kimi memberships and Kimi Code subscriptions have their own usage arrangements. These token rates apply to pay-as-you-go API requests.

A budget of 100,000 fresh input tokens and 10,000 output tokens is $0.30 on Pro or $0.60 on Highspeed before separate cache-write charges. Across a tool loop, add the billable usage of every turn, including reasoning output and retries. A one-million-token context limit is capacity, not an included allowance.

Kimi automatically caches eligible repeated prefixes, with a default five-minute lifetime. Writes have a separate charge determined by the 5m or 1h tier. A cache hit is charged at the cache-read rate without another write charge for that hit. An expired or changed prefix can require another write.

For example, a later request reading one million already-cached tokens and producing 10,000 output tokens costs $0.30 on Pro or $0.40 on Highspeed. The earlier write has its own bill: $2 or $4 for a one-million-token Pro prefix at 5m or 1h, and $4 or $8 for Highspeed. Add fresh input separately. Spread creation costs over the actual reuses when evaluating savings.

Put stable instructions, tools and reference files first, with changing questions and tool results afterward. Select prompt_cache_options.ttl="1h" for reuse intervals beyond five minutes, then inspect the usage record to see what was read and written. Repeated wording alone does not establish a cache hit.

The international platform requires a successful top-up of at least $1 to unlock this flagship model. Create an API key, install the OpenAI Python SDK and set MOONSHOT_API_KEY on your server. Use https://api.moonshot.ai/v1 for the USD route below. The returned message contains the answer; this single request does not execute a coding agent’s tools.

Set model to kimi-k3.1-pro-highspeed for the faster variant. Thinking is always enabled; reasoning_effort supports low, high and max, with max as the default. Preserve the complete returned assistant message when continuing a tool conversation, including reasoning and tool calls.

Moonshot describes Highspeed as two to three times faster than Pro. Measure end-to-end latency on your own workload: tool execution, prompt processing and retries also affect total time.

For interactive work where waiting is expensive, compare Pro and Highspeed on the same accepted result, effort level and cache state. For background batches with a flexible deadline, start with standard Pro’s lower fresh-input and output rates. Cache-heavy tasks can have a smaller price gap because cache reads cost the same on both variants.

Atomic Agent authors Nadya Dudka and Andrew Dyuzhov tested standard Pro through the direct Kimi API on October 23, using Atomic Agent 0.6.9 and temperature 1. Their Python package had five planted bugs, two missing features and 32 hidden tests, repeated three times per effort setting: 96 test checks per setting, rather than 96 different tasks. They report average completion times of 2.2 minutes at low effort and 4.9 minutes at max, with estimated task-cost ranges of $0.16–$0.21 and $0.35–$0.46 respectively.

The ranges reflect an unresolved billing difference: their account balance fell $2.02 across all six Kimi runs, while their log-based calculation was $1.52. They could not assign the difference to individual effort settings. About 96% of input was cached. This small, cache-heavy workload illustrates an effort trade-off; it does not price Highspeed, subscriptions or a different repository task.

Moonshot’s current announcement commits to releasing full weights by November 4. It describes a 3T-class model; this is insufficient to choose a quantized checkpoint, serving recipe or GPU allocation before the files and runtime requirements are available. Budget the hosted route for current work.

Kimi K3 remains a separate model. A K3 hardware configuration or price does not establish a K3.1 Pro deployment. Compare MiniMax M3.1 Flash if downloadable weights and documented server recipes are part of the decision today.