Kimi K3.1 Pro API Pricing: $2/M Input Tokens

Kimi K3.1 Pro API pricing is $2.00 per million input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens on Moonshot's price page (checked November 1, 2026). In nine real requests, one call cost $0.0026 to $0.0556, and output tokens, most of them reasoning, were the larger part of the bill in eight of the nine.

Short answer

• $2.00 in, $10.00 out, $0.20 cached. The rate is the same at every prompt length, up to the full 1,048,576-token context (Moonshot's K3.1 Pro quickstart, checked November 1, 2026).
• Output carries the bill. Thinking is billed as output, and output was more than half the cost in eight of our nine requests, 84% to 97% of it in seven.
• Caching cut input cost 88%. We asked three questions about the same 7,300-token catalog: input cost $0.01468 on the first call and about $0.00175 on each of the next two.
• Set the effort level yourself. High effort cost 43% more than low on the same explanation and 82% more on the same code task.

How much does Kimi K3.1 Pro cost?

Every figure is USD per million tokens, from Moonshot's price page (checked November 1, 2026). Highspeed doubles input and output while cached input stays at $0.20. Kimi K3, the previous flagship, sits between the two K3.1 tiers.

The rate is flat at every length: a prompt near the 1,048,576-token limit pays the same $2.00 per million as a short one, and output caps default to 131,072 and can be raised to 1,048,576 (quickstart, checked November 1, 2026).

A cached input token costs $0.20 instead of $2.00, a 90% discount, and each hit refreshes the entry's time to live; 5 minutes is the default. A 5-minute write costs $2.00 per million, the same as fresh input, and a 1-hour write costs $4.00 per million.

Batch runs at 60% of list, $1.20 input and $6.00 output, for kimi-k3.1-pro and Kimi K3, and still pays cache-write fees with best-effort cache hits (Batch price page, checked November 1, 2026). On Moonshot's China platform the same model is ¥13.00 in, ¥68.00 out and ¥1.30 cached.

Tools bill per call: $0.002 for Web Search Basic, $0.003 for Web Search Pro, $0.002 for URL Fetch and $0.005 for the Responses API web_search tool, with failed or empty searches free (tool pricing, checked November 1, 2026). Prices exclude taxes.

How is one request billed?

Three lines decide a bill: uncached input, cached input and output.

cost = uncached input / 1M × $2.00
     + cached input    / 1M × $0.20
     + output          / 1M × $10.00

Reasoning sits inside the output line. Moonshot's thinking-model guide says reasoning_content counts toward your input/output token quota. You cannot switch thinking off either; asked how, Moonshot's quickstart answers: "You can't — K3.1 Pro always thinks. If the reasoning takes too long, set reasoning_effort to low" (checked November 1, 2026).

One explanation request from our test set: 117 fresh input tokens, 470 output tokens, 298 of them reasoning. Input cost 117 / 1M × $2.00 = $0.00023, output 470 / 1M × $10.00 = $0.00470, so the call cost $0.00493, 95.3% of it output.

What did nine real requests cost?

We ran nine requests on November 1, 2026 with kimi-k3.1-pro through OmniaKey's OpenAI-compatible API, streaming, with max_tokens at 4,000 for the catalog questions, 8,000 for the explanations and 32,000 for the code. The three catalog questions re-sent the same 450-row CSV, about 7,300 tokens. Costs come from the API's usage field at Moonshot's list rates rather than OmniaKey's: $2.00 input, $0.20 cached, $10.00 output.

The nine calls cost $0.09989 in total, from $0.00261 to $0.05564. The CSV questions were lookups, the explanation asked for about 120 words on prompt caching, and the code asked for a merge_bookings function checked by 10 unit tests.

Did prompt caching help?

Yes, on the two calls that hit the cache. In questions 2 and 3, 7,168 of about 7,320 input tokens came from cache and the input line fell from $0.01465 to $0.00175, an 88% cut; question 3 cost $0.00261 instead of $0.01551. We sent no cache settings; the hits showed up in the usage OmniaKey's API returned, with 0 cache-write tokens. 10,000 lookups like question 3 cost about $26.08 with the cache and $155.10 without it.

How much does reasoning_effort change the bill?

High effort cost 43% more than low on the explanation and 82% more on the code, because it wrote 45% and 96% more output: 680 against 470, and 624 against 318. All three code runs passed 10 of 10 tests.

Moonshot's reasoning-effort guide lists low, high and max, with max as the default (checked November 1, 2026). In our runs the default behaved like low: 472 against 470 output tokens on the explanation, 310 against 318 on the code. That is one run per setting, so set the level and measure. Low has its own cost: it answered 25 where the correct catalog count was 26.

Where do people overspend?

• A cap that reasoning swallows. With max_tokens at 4,000 and low effort, the first catalog request stopped at the cap. The API reported 4,096 output tokens, 4,093 of them reasoning, and an empty answer, and the call still cost $0.05564, $0.04096 of it output. Size max_completion_tokens above the answer plus its thinking.
• The default effort level. Moonshot's docs name max as the default. Hacker News commenter willmorrison wrote on October 23, 2026 about the $100/month subscription, not the API: "Bought the $100/month subscription and hit my 5-hour limit with one prompt on default settings (max effort) in their CLI harness."
• Highspeed at double the rate. The same 117-in, 470-out explanation costs $0.00987 on kimi-k3.1-pro-highspeed instead of $0.00493.
• Verbosity. Artificial Analysis counted 110M tokens for K3.1 Pro across its Intelligence Index evaluations, against a median of 70M, and prices one index task at $3.68 on average from the tokens each task used (checked November 1, 2026).

How does it compare with Claude Sonnet 5.5, Opus 5.5, GPT-6 Astra and GLM-5.3?

Kimi K3.1 Pro has the same input and output prices as Claude Sonnet 5.5, yet the two sit far apart on the cost of a finished task.

Token prices come from Anthropic, OpenAI, Z.ai and DeepSeek, all checked November 1, 2026; DeepSeek V4 Pro shows peak-hour rates, with off-peak at half. The last column is Artificial Analysis's average cost per Intelligence Index task (checked November 1, 2026), which its page leaves out for DeepSeek V4 Pro.

At the same $2.00 and $10.00, Kimi K3.1 Pro costs $3.68 per index task against Sonnet 5.5's $0.78, 4.7 times as much. One user's own benchmark went the other way: thefourthchime wrote on Hacker News on October 23, 2026 that "Kimi K3.1 Pro ties Sonnet 5.5 at 93, for $0.68 and 11 min vs Sonnet's $1.61 and 14 min." Your task decides which pattern you get, so price a few of your own first; our K3.1 Pro review has the benchmark scores.

Current rates for the five rivals are on the Claude Sonnet 5.5, Claude Opus 5.5, GPT-6 Astra, GLM-5.3 and DeepSeek V4 Pro model pages.

DeepSeek's peak and off-peak rates are explained in DeepSeek V4 Pro API pricing.

How do you cut the bill?

1. Set reasoning_effort yourself: the default is max, high effort cost 43% to 82% more than low in our runs, and low passed all 10 code tests.
2. Leave output headroom for thinking: reasoning bills as output and can fill the cap, as that 4,093-token first run showed.
3. Keep the stable part of the prompt first and identical between calls: our CSV came back three times, and the last two calls read 7,168 tokens from cache at $0.20 instead of $2.00.
4. Send offline jobs through Batch at 60% of list, $1.20 in and $6.00 out per million; the explanation would cost about $0.00296, and batch cache hits are best effort.
5. Stay on the standard tier unless waiting costs more than tokens; Highspeed doubles both rates.

Kimi K3.1 Pro is also on OmniaKey's OpenAI-compatible API, the same API our nine runs used, so one key covers it and the five rivals above. Current rates are on the Kimi K3.1 Pro model page and the pricing page.

Frequently asked questions

How much does Kimi K3.1 Pro cost per million tokens?

$2.00 for input, $10.00 for output and $0.20 for cached input, per million tokens (Moonshot's price page, checked November 1, 2026). Highspeed doubles input and output, and cached input stays at $0.20.

Are Kimi K3.1 Pro reasoning tokens billed?

Yes. Moonshot's thinking-model guide says reasoning_content counts toward your input/output token quota, so reasoning bills at the $10.00 output rate. In our runs it ranged from 15.8% to 99.9% of output tokens.

Can you turn off thinking on Kimi K3.1 Pro?

No. Moonshot's quickstart says K3.1 Pro always thinks and suggests setting reasoning_effort to low when reasoning takes too long. Low, high and max are the three levels, with max as the default.

How much do you need to top up to use Kimi K3.1 Pro?

$1. Moonshot's limits page (checked November 1, 2026) says a $1 minimum top-up unlocks the model, with 1 concurrent request, 3 requests per minute, 500,000 tokens per minute and 1,500,000 tokens a day. A cumulative $5 top-up brings a $5 voucher, and $10 cumulative raises the limit to 100 requests per minute with no daily cap; vouchers do not count toward the total.

What happens to Kimi K2 pricing?

K2 models closed to new users on October 31, 2026, existing users keep access, and every K2 model retires by the end of December 2026 (Moonshot's price page, checked November 1, 2026). Kimi K2.7 Code (kimi-k2.7-code) lists $0.95 input, $4.00 output and $0.19 cached input per million, with a 262,144-token context.

Sources checked

• Moonshot: chat model pricing
• Moonshot: quickstart
• Moonshot: reasoning effort
• Moonshot: thinking models
• Moonshot: rate limits
• Moonshot: Batch pricing
• Moonshot: tool pricing
• Moonshot: China platform pricing
• Anthropic: API pricing
• OpenAI: API pricing
• Z.ai: model pricing
• DeepSeek: API pricing
• Artificial Analysis: K3.1 Pro
• Hacker News: thefourthchime on K3.1 Pro and Sonnet 5.5
• Hacker News: willmorrison on the default effort