Kimi K3 API Pricing 2026: K3, K3.1 Pro and Every Kimi Model per Million Tokens, Cache Writes, Limits and Cost per Conversation

Published 28 October 2026. Every Moonshot price below was read from the Kimi API Platform pricing page and its Chinese counterpart on 28 October 2026, and every price is as of that date.

Kimi K3 costs $3.00 per million input tokens, $0.30 per million cached input tokens and $15.00 per million output tokens on Moonshot AI's own API, as of 28 October 2026, with one flat rate across its 1,048,576-token context window. It is no longer the model to price first. Kimi K3.1 Pro, which Moonshot released on 24 October as a post-trained version of K3, costs $2.00, $0.20 and $10.00, a third less on every line, and Moonshot's own billing FAQ now says "If cost is a concern, you can use the kimi-k3.1-pro model." Both models always reason, both bill that reasoning as output, and both now charge cache writes as a separate line: a 5-minute write at the input rate and a 1-hour write at twice it. The API is prepaid with no free tier, and a $1 top-up opens K3 and K3.1 Pro at the lowest rate-limit tier.

Kimi K3 itself launched on 16 July 2026, according to Moonshot's blog index, as a 2.8-trillion-parameter open-weight model with native vision, and many of the pages ranking for "kimi k3 pricing" were written that week. Three things have changed since. K3.1 Pro arrived at a lower price than K3. Moonshot split cache writes out of the input price into the 5-minute and 1-hour tiers, so guides that say Kimi has "no cache-write premium" are now half right. And the K2 models those guides list as the cheap alternative stop accepting newly registered users on 31 October 2026 and retire by the end of December. In the worked example further down, a 40-turn conversation costs $1.64 on K3 and $1.09 on K3.1 Pro without caching, and $0.411 and $0.274 with it; K3.1 Pro's uncached figure is the same as Claude Sonnet 5.5's.

Kimi API price per million tokens, every current model

US dollars per million tokens, excluding tax, from the Kimi API Platform pricing page on 28 October 2026. Both platform.moonshot.ai and platform.moonshot.cn now redirect to Kimi-branded domains, platform.kimi.ai and platform.kimi.com, and the price tables on those pages are drawn in by script; the Markdown copies Moonshot publishes for AI tools show the same tables empty.

The order of the flagship rows is the surprise. K3.1 Pro is the newer and, by Moonshot's account, the more capable model, and it costs two-thirds of K3 on input, cached input and output. Moonshot describes K3 on its model list as "Kimi's previous-generation flagship", so there is little reason to start a new project on kimi-k3 unless you are pinning behaviour you have already tested. The Highspeed variant of K3.1 Pro, which Moonshot says runs "about 2-3× the speed", doubles the input and output price but keeps the $0.20 cache-hit rate, so on a heavily cached agent loop it costs much less than double.

The K2 rows come with a deadline. The pricing page says that "starting October 31, 2026, K2 series models will no longer be available to newly registered users", existing users keep access, and Moonshot plans "to retire all K2 series models by the end of December 2026". An account created on 1 November can call only the K3 series. Older names are already gone: kimi-k2.5 and every moonshot-v1 model were discontinued on 31 August 2026, the original kimi-k2 series on 25 May 2026 and kimi-latest on 28 January 2026, according to the model list, and calls to them return a 404.

Kimi API prices in CNY on platform.kimi.com

Moonshot runs two platforms with separate price lists, and neither is a currency conversion of the other. The global platform at platform.kimi.ai bills in US dollars, under terms with Moonshot AI PTE. LTD. and Singapore law; the China platform at platform.kimi.com bills in renminbi.

Renminbi per million tokens from platform.kimi.com, read 28 October 2026. The ratios differ between the two lists: K3.1 Pro's output is five times its input in dollars and roughly 5.2 times in renminbi, and the K2 cache-hit rates are not the same fraction of input on both. Budget in the currency you will be billed in, and do not convert one list into the other. The rate-limit thresholds on the China platform are also in renminbi (Tier 1 at ¥50 of cumulative top-ups, against $10 on the global platform), and the China platform's Tier 0 starts at ¥0.

How Kimi caching is billed: automatic, with 5-minute and 1-hour writes

Caching on the Kimi API is on by default. Moonshot "automatically enables context caching billing for repeated request prefixes, with a default TTL (time-to-live) of 5m", according to the pricing page, and its context caching guide lays out the billing for the K3 series. Every input token in a request falls into exactly one of three buckets, which "are mutually exclusive and sum to total input tokens": a cache hit billed at the cached rate, a cache write billed at the write rate for its TTL, or an uncached remainder billed at the input rate.

On kimi-k3 that means a hit costs $0.30 per million, a 5-minute write costs $3.00 (the same as an uncached token) and a 1-hour write costs $6.00. The guide is explicit that "the write cost was already included in the input price, so existing requests cost the same overall"; the split makes the write visible on the bill rather than adding a charge to the default. The 1-hour TTL is the only option that costs more up front, and Moonshot's arithmetic is that two hits within the hour repay the extra write fee. A hit renews the entry under its original TTL at no charge, a TTL cannot be changed until the entry expires, and the 5-minute and 1-hour caches do not share entries.

You choose the TTL with prompt_cache_options: {"mode": "implicit", "ttl": "1h"} on the Chat Completions and Responses APIs, or a top-level cache_control with a ttl on the Anthropic-compatible Messages API. A few details catch people out. Cache is stored in blocks, and "a portion smaller than one full block cannot be written to cache" and bills at the input rate; Moonshot does not publish the block size. Cache is scoped to an organization, shared across its projects and never across organizations. And only kimi-k3.1-pro, kimi-k3.1-pro-highspeed and kimi-k3 support cache writes; the K2 models bill hits at their cached rate with no write line. Usage fields report the split: usage.prompt_tokens_details.cached_tokens and cache_write_tokens on Chat Completions, with response headers Msh-Usage-Cache-Write-Tokens-5m and -1h showing which TTL was written.

For a chat product the practical rule follows from the default. A user who replies within five minutes keeps the cache warm for free, because each hit renews it; a user who pauses for ten minutes and comes back pays to write the prefix again. Set ttl: "1h" on conversations where people stop to think, and leave the default on agent loops that call the model continuously.

Free tier, top-ups and rate limits

The Kimi API has no free tier and no subscription. Access is prepaid: Moonshot's recharge and rate limits page says "a minimum recharge of $1 is required to start using the API; when your cumulative recharge reaches $5, you receive a $5 voucher", and the K3 and K3.1 Pro quickstarts both describe the models as "unlocked after a successful top-up (minimum $1)". Vouchers do not count towards your tier, and since September 2026, according to the platform changelog, usage is deducted "at a uniform ratio of 50% cash and 50% vouchers" until either balance runs out, so a voucher stretches your cash rather than being spent first.

Your rate limits follow your cumulative top-ups, not your spend:

Source: Moonshot's recharge and rate limits pages on platform.kimi.ai and platform.kimi.com, read 28 October 2026.

Tier 0 is a smoke test: three requests a minute, 500,000 tokens a minute and 1.5 million tokens a day, so a single prompt longer than 500,000 tokens cannot run until Tier 1, and the 1M-token context is out of reach on a $1 account. The step from $1 to $10 matters most, because it lifts concurrency from 1 to 15, requests from 3 to 100 a minute and removes the daily token cap. Moonshot also warns that limits "may be temporarily adjusted" when its cluster reaches capacity, and that a risk-control limit triggered by abnormal activity "cannot be lifted". Above Tier 5, the route is the rate-limit increase form or an enterprise contract.

Spending controls sit on projects. Moonshot's billing FAQ recommends a project daily budget, after which "the system will automatically reject all API requests under that project", with the caveat that "due to billing latency, the limit may take about 10 minutes to take effect"; daily and monthly project limits can also be set through the Admin API. On a $15-per-million-output model running max reasoning effort, ten minutes is long enough to matter, so set the budget below what you can afford to lose.

Reasoning tokens: always on, billed as output, and sent back as input

K3 and K3.1 Pro do not have a non-thinking mode. The K3 quickstart says "K3 always has thinking mode enabled" and that reasoning_effort accepts low, high and max, with max the default; K3.1 Pro behaves the same way. The thinking arrives in reasoning_content, counts against max_completion_tokens (default 131,072, settable up to 1,048,576) and bills at the output rate, which is $15 per million on K3 and $10 on K3.1 Pro.

The multi-turn rule doubles the effect. Moonshot's thinking models guide says that for K3.1 Pro and K3, passing the previous assistant message back with its reasoning_content "is required in multi-turn conversations and tool-call loops", and that "historical thinking content keeps occupying the context window and is billed accordingly". Each turn's reasoning is paid once as output and again as input on every later turn. K2.6 differs: it drops historical thinking by default (thinking.keep is null), while K2.7 Code always keeps it.

To see what that does, run the conversation below with 1,000 reasoning tokens on every reply, all replayed. The 40th request then carries 60,650 input tokens instead of 21,650, and the 40-turn bill on K3 rises from $1.64 to $4.58 without caching and from $0.411 to $1.35 with it; on K3.1 Pro it rises from $1.09 to $3.05 and from $0.274 to $0.900. The 1,000-token figure is an illustration, not a measurement, and at the default max effort real traces can be longer. Setting reasoning_effort: "low" on routing, extraction and short-answer turns is the largest single lever on a Kimi bill after choosing the model.

What a 40-turn conversation costs on Kimi, DeepSeek, Mistral, Groq, Gemini and Claude

Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens; these are the same assumptions as our Groq API pricing, DeepSeek API pricing and Claude API pricing guides, so the figures line up. Turn 1 sends 2,150 input tokens and turn 40 sends 21,650, because the API is stateless and each request carries every exchange before it. The whole 40-turn conversation consumes 476,000 input tokens to produce 14,000 output tokens, and a 100-turn conversation consumes 2,690,000 input tokens and 35,000 output tokens. Reasoning tokens are left out of the table, which flatters K3 and K3.1 Pro more than any other model in it, since they cannot turn reasoning off.

Kimi rows are our arithmetic on Moonshot's US-dollar rates read 28 October 2026. The other rows are the figures our sibling guides published: DeepSeek, Groq, Gemini and Claude from rates read on 27 October 2026, and Mistral from rates read on 28 October 2026; the method for each is on its own page. In renminbi on platform.kimi.com, the same conversation costs ¥10.92 uncached and ¥2.74 cached on K3, and ¥7.14 and ¥1.82 on K3.1 Pro.

"Cache warm" on Kimi means that from turn 2, the previous request's prompt is read at the cached rate, and the 500 new tokens (the last reply plus the new message) are written to the 5-minute cache at the input rate, which on the K3 series costs the same as an uncached token. Turn 1 is a write of the full 2,150 tokens. Moonshot does not publish its block size, so the small remainder that misses a block and bills at the full rate is left out, which makes the warm figures a slight underestimate. The model assumes every reply comes within five minutes; a user who pauses longer forces a rewrite of the prefix.

Read across, the table makes the pricing position plain. Without caching, K3.1 Pro costs exactly what Claude Sonnet 5.5 and Gemini 4 Argon (until 30 November) cost, $1.09 for 40 turns, because all three list at $2 input and $10 output; K3 at $1.64 costs half as much again, and the Highspeed variant matches Claude Opus 5.5 at $2.18. With caching, K3.1 Pro at $0.274 costs slightly more than Sonnet 5.5 at $0.240, because a Kimi cache hit costs $0.20 against Claude's $0.10, even though Claude charges 1.25 times input to write its cache and Kimi charges nothing extra. And against the low-cost models the gap is an order of magnitude: DeepSeek deepseek-flash off-peak, Mistral Small 4 and Groq GPT-OSS 120B all cost $0.080 uncached, a fourteenth of K3.1 Pro. Kimi K3 is priced as a frontier model, not as a cheap open-weight alternative, and the K2 models that were the cheap option close to new accounts on 31 October.

At 10,000 forty-turn conversations a month, K3.1 Pro costs $10,920 without caching and $2,742 with it, K3 costs $16,380 and $4,113, and K2.6, for an existing account, costs $5,082 and $1,493. On Claude Sonnet 5.5 the same volume costs roughly $10,920 and $2,396, according to our Claude guide. Add 1,000 replayed reasoning tokens per reply and the K3.1 Pro figures roughly triple.

The table also shows the shape of the problem. Going from 40 to 100 turns multiplies turns by 2.5 and the uncached K3.1 Pro bill by 5.3, because every turn carries every turn before it, and Kimi's required reasoning replay steepens the curve further. Caching cuts the price of the repeated tokens; it does not stop them being sent, and a 5-minute or 1-hour cache does nothing for a user who comes back tomorrow. The other lever is to stop resending the whole transcript. We build Maximem Synap, a memory layer that returns a ranked set of facts about the user, trimmed to a token budget you set, in place of the full history. When that beats caching, and when it does not, is worked through in how to reduce LLM token costs in long conversations.

What third-party hosts charge for Kimi K3

K3 is an open-weight model, and Moonshot says the full weights have been released, so other inference providers serve it at their own prices. On 28 October OpenRouter's endpoint data listed 21 providers for K3, while K3.1 Pro, whose weights Moonshot says it will release "by Nov 4", was served only by Moonshot.

Prices per million tokens in US dollars, read from each provider's page on 28 October 2026; OpenRouter figures from its public endpoints API on the same day.

Together and Fireworks match Moonshot's list price exactly, so the reason to use them is operational: a contract you already have, a US-only endpoint on Fireworks, or zero data retention without an enterprise agreement. DeepInfra is 5% cheaper on every line but charges a cache-write premium that Moonshot does not, so on a chat workload the saving shrinks. Groq, which served Kimi K2 earlier this year, does not offer K3; our Groq pricing guide records Kimi K2's removal from its plans.

The low OpenRouter numbers need reading carefully. The cheapest K3 input, $0.15, comes from an endpoint that charges $20 for output, and the cheapest output, $8.25, includes a 45% provider discount shown in the same data. Many of the cheaper listings serve 4-bit quantized weights (fp4 or mxfp4), according to OpenRouter's endpoint data, which is a different trade from Moonshot's own endpoint. Aggregator pages that headline "Kimi K3 from $0.19 per million tokens" are quoting one of these, not Moonshot's rate. A cached prefix also only pays off when consecutive requests land on the same provider, so if you route through OpenRouter, pin a conversation to one provider or the cache rebuilds on every hop.

Batch, web search and file pricing

Moonshot's Batch API costs 60% of the standard price, according to its Batch pricing page, but it supports only kimi-k2.7-code and kimi-k2.6: $0.57 input, $2.40 output and $0.114 or $0.10 cached per million tokens. K3 and K3.1 Pro have no batch rate on Moonshot. Batch jobs also fall outside zero data retention, and with K2 closing to new users on 31 October, batch on Moonshot is effectively an existing-customer feature. If you need K3 at a batch discount, OpenRouter listed a DeepInfra batch endpoint for K3 at $2.28 input and $11.40 output on 28 October, with a 16,384-token output cap.

The web tools bill per call, on Moonshot's WebSearch pricing page: $0.002 per successful call to Web Search Basic (/v1/tools/search), $0.003 to Web Search Pro, $0.002 to URL Fetch, and $0.005 per executed search through the Responses API web_search tool or the Formula web-search tool. A call is billed only when it returns non-empty results, and the search results you pass back to the model bill as ordinary input tokens. Search endpoints have their own QPS limit (1 at Tier 0 up to 50 at Tier 5) that does not draw on your RPM or TPM.

File extraction and file storage are "temporarily free", according to the pricing page; the extracted text you then send to the model bills as input.

Kimi membership plans are not API credit

A Kimi membership and the Kimi API are separate products with separate bills. Moonshot's product comparison says the API "uses pay-as-you-go billing with no subscription plan", and that Kimi Code, a subscription coding assistant, has an API that "is independent from the API service on this platform, and the two are not interchangeable".

For reference, Moonshot's own Kimi K3 pricing page, updated 27 October, lists memberships at $0 (Free), $15, $31, $79 and $159 a month on annual billing for Plus, Pro, Max and Ultra, or $19, $39, $99 and $199 billed monthly, with only Max and Ultra opening K3's 1M-token chat context. The page lists no API credit with any of them. That same page lists K3's API price but not K3.1 Pro or the cache-write tiers.

Data use, zero data retention and the contract

Moonshot's Terms of Service, last updated 30 July 2026, are with Moonshot AI PTE. LTD. under Singapore law, and section 4 says Moonshot "may use Content to provide, maintain, develop, support, and improve the Services", with customers who need restrictions on training told to "contact Moonshot AI to discuss available enterprise arrangements". Moonshot's zero data retention page adds that "by default, the Kimi Open Platform does not use enterprise customer data to train models" and offers ZDR "on request to enterprise customers". Read together, a pay-as-you-go account without a written agreement should assume its content can be used to improve the service.

ZDR has exclusions: Batch jobs, images and video uploaded through the Files API, operational data such as billing records, and third-party tools. Fireworks states that zero data retention is on by default for its Kimi K3 endpoint, and DeepInfra marks its endpoint "zero retention", which makes a third-party host the faster route to a no-retention setup for a team without an enterprise contract with Moonshot.

Which Kimi model to pick on cost

For new work on Moonshot's API, start with K3.1 Pro. It is the current flagship, it costs $2.00 input and $10.00 output against K3's $3.00 and $15.00, it has the same 1M-token context and cache-write options, and Moonshot itself points cost-conscious users to it. Keep kimi-k3 only for workloads you have already evaluated on it, and plan the move. Use K3.1 Pro Highspeed where latency is the product, since its doubled input and output prices come with the same $0.20 cache-hit rate.

If you need K3 specifically and want a second supplier, Together and Fireworks sell it at Moonshot's list price and DeepInfra at 5% less, and the open weights leave self-hosting open, though a 2.8-trillion-parameter model is a large serving bill of its own. Existing accounts on K2.6 or K2.7 Code can keep using them at $0.95 and $4.00 until the end-of-year retirement, but a new account cannot choose them after 31 October, so they are not a plan for anything that ships next year.

If the workload is high-volume chat where the model does not need to be frontier-class, the comparison table is the honest answer: K3.1 Pro costs about fourteen times what DeepSeek deepseek-flash, Mistral Small 4 or Groq's GPT-OSS 120B cost for the same 40-turn conversation, before its reasoning tokens.

How to estimate your monthly Kimi bill, and cut it

A K3-series request costs (cache-hit tokens times the cached rate) plus (cache-write tokens times the write rate for their TTL) plus (uncached input tokens times the input rate) plus (output and reasoning tokens times the output rate). For a conversation, the input on each turn is the system prompt plus every earlier turn, including replayed reasoning_content, so estimate per conversation as in the table above and multiply by conversations per month. Add web search calls at $0.002 to $0.005 each, and tax at checkout.

The levers that move a Kimi bill, roughly in order of size:

1. Use K3.1 Pro instead of K3. It is a third cheaper on every token.
2. Lower reasoning_effort where the answer is simple. Reasoning bills as output and returns as input on every later turn.
3. Keep the prefix stable and pick the right TTL. Put the system prompt and tool definitions first, keep timestamps out of them, and use the 1-hour cache when users pause.
4. Cap max_completion_tokens. The 131,072-token default lets a single runaway reply cost up to $1.97 on K3.
5. Set a project daily budget, remembering the roughly ten-minute enforcement lag.
6. Stop resending the whole conversation once chats run long, by summarising or retrieving only what the next turn needs.

Frequently asked questions

How much does the Kimi K3 API cost?

As of 28 October 2026, Kimi K3 costs $3.00 per million input tokens, $0.30 per million cached input tokens and $15.00 per million output tokens on Moonshot's API, with a 5-minute cache write at $3.00 and a 1-hour cache write at $6.00. The price is flat across the full 1,048,576-token context window.

How much does Kimi K3.1 Pro cost?

Kimi K3.1 Pro costs $2.00 per million input tokens, $0.20 cached and $10.00 output, with cache writes at $2.00 for 5 minutes and $4.00 for 1 hour. The Highspeed variant costs $4.00 input and $20.00 output with the same $0.20 cached rate. K3.1 Pro was released on 24 October 2026 and is cheaper than K3.

Is there a free tier for the Kimi API?

The Kimi API has no free tier: it is prepaid with a $1 minimum top-up, which opens K3 and K3.1 Pro at Tier 0 (3 requests a minute, 1.5 million tokens a day). Reaching $5 in cumulative top-ups earns a $5 voucher, and usage is drawn half from cash and half from vouchers.

What are the Kimi API rate limits?

Limits depend on cumulative top-ups: Tier 0 ($1) allows 1 concurrent request, 3 RPM and 500,000 TPM; Tier 1 ($10) allows 15, 100 RPM and 2,000,000 TPM with no daily cap; Tier 5 ($3,000) allows 100, 300 RPM and 5,000,000 TPM.

What is the Kimi K3 context window?

Kimi K3 and K3.1 Pro have a 1,048,576-token context window, and output (max_completion_tokens) defaults to 131,072 tokens and can be set up to 1,048,576. The K2.6 and K2.7 Code models have a 262,144-token window.

Does Kimi K3 charge for reasoning tokens?

K3 and K3.1 Pro always reason, the reasoning bills at the output rate, and Moonshot requires you to send previous reasoning_content back in multi-turn conversations, so it bills again as input on later turns. reasoning_effort accepts low, high and max, with max the default.

What is Moonshot API pricing in CNY?

On platform.kimi.com, K3 costs ¥20 input, ¥2 cached and ¥100 output per million tokens, and K3.1 Pro costs ¥13, ¥1.30 and ¥68. The China platform has its own price list rather than a conversion of the US-dollar rates.

How much does Kimi K3 cost on OpenRouter, Together or Fireworks?

Together AI and Fireworks list Kimi K3 at $3.00 input, $0.30 cached and $15.00 output, the same as Moonshot; DeepInfra lists $2.85, $0.285 and $14.25. OpenRouter routes K3 across 21 providers, with some quantized endpoints much cheaper on one line and more expensive on the other. Groq does not serve K3.

Does the Kimi Batch API support K3?

Moonshot's Batch API does not support K3; at 60% of the standard price, it supports only kimi-k2.7-code and kimi-k2.6 as of 28 October 2026.

Can new users still use Kimi K2.6?

New accounts cannot use K2.6 after 31 October 2026, when the K2 series closes to newly registered users; existing users keep access, and Moonshot plans to retire all K2 models by the end of December 2026.

Is Kimi K3 cheaper than Claude, Gemini or DeepSeek?

Kimi K3 is priced as a frontier model, not as a low-cost one. A 40-turn conversation costs $1.64 on K3 and $1.09 on K3.1 Pro without caching, against $1.09 on Claude Sonnet 5.5 and Gemini 4 Argon and $0.080 on DeepSeek deepseek-flash off-peak. With caching, K3.1 Pro costs $0.274 and Sonnet 5.5 costs $0.240.

Sources

Moonshot and host sources retrieved 28 October 2026; comparison figures from our guides as dated in the table.

• Moonshot (Kimi API Platform, global): model inference pricing, Batch pricing, WebSearch pricing, recharge and rate limits, model list, Kimi K3 quickstart, Kimi K3.1 Pro quickstart, context caching, thinking models, account and billing, product comparison, zero data retention, Terms of Service and platform changelog
• Moonshot (China platform): pricing in CNY, rate limits in CNY and Batch pricing in CNY
• Kimi: Kimi K3 tech blog, Kimi K3.1 Pro tech blog, blog index with release dates and Kimi K3 pricing page
• Hosts: Together AI Kimi K3 and pricing, Fireworks Kimi K3, DeepInfra Kimi K3, OpenRouter Kimi K3 and its public models API, and Groq models
• Maximem: Groq API pricing, DeepSeek API pricing and Claude API pricing

Worked examples are our own arithmetic on those rates.