GenGrades

Kimi K3.1 Pro review: pricing, API and vs Claude Opus 5.5

Published Updated

Moonshot AIKimi K3.1 Pro logoAPI availableOpen weightsLicense: Kimi K3.1-Pro License (modified MIT)

What Kimi K3.1 Pro is

In one or two sentences, before anything else.

Kimi K3.1 Pro is Moonshot AI's open-weight 2.8-trillion-parameter model with a 1M-token context, released October 23, 2026, weights out November 4. Its API costs 2 and 10 USD per million tokens, half Claude Opus 5.5's. Vals AI puts it near Opus on Terminal-Bench 4.0, but independent aggregates rank it lower, it is slow, and an audit questions its benchmark lead.

Main features

What it can do, one sentence each, and whether we actually tested it.

  • 1M-token context at one flat price

    Reads up to 1,048,576 tokens per request and writes up to 131,072 by default, or up to 1,048,576 when raised, with no surcharge for long prompts.

    Per Moonshot's API quickstart and price list. Context caching is automatic, with 5-minute and 1-hour cache tiers billed separately.

  • Reasoning that is always on

    Thinks before every answer; each request picks low, high or max reasoning effort, and max is the default on the API.

    Per Moonshot's docs the chain of thought cannot be switched off. Artificial Analysis counted 113 million output tokens to run its index, above the median of 104 million for comparable models.

  • Image and video input

    Reads photos, charts, documents and, through uploaded files, video, using a 401-million-parameter MoonViT-V2 vision encoder; output is text.

    Per Moonshot's model card and quickstart. The API does not accept public image URLs: images go in as base64 or uploaded files.

  • Highspeed variant

    Serves the same model two to three times faster as kimi-k3.1-pro-highspeed, at 4 USD input and 20 USD output per million tokens.

    Per Moonshot's price list and Kimi Code docs, Highspeed costs twice as much on the API and uses about twice the quota in Kimi Code; the Kimi app offers it from the Pro plan up.

  • OpenAI- and Anthropic-compatible API

    Accepts OpenAI Chat Completions and Responses requests and Anthropic Messages requests, so Claude Code, Codex and existing SDK code can switch by changing the base URL and key.

    Per Moonshot's API overview. Its Kimi Code docs warn that translating protocols through proxy tools can change the model's behaviour.

  • Open weights for self-hosting

    Publishes the full 2.8T-parameter checkpoint on Hugging Face, about 1.56 TB, trained to run in 4-bit MXFP4, ungated and with day-0 support in vLLM and SGLang.

    Per the vLLM recipe, which sizes the MXFP4 variant at 1,874 GB, it runs on one 8-GPU B300 node, two GB300 trays, one AMD MI355X node or a 16-chip TPU v7x slice.

  • Agentic coding and 3D scripting

    Works as a terminal coding agent and writes scripts for Blender, FreeCAD and game engines such as Unity, Unreal and Godot to build 3D assets and scenes.

    Moonshot reports 65.7% on Terminal-Bench 4.0 and 92.7% on BenchCAD with tools; Vals AI's own Terminal-Bench run gave 63.64%, and DesignArena voters rank it fourth on 3D design.

  • Request signatures

    Returns a signature with each API response that anyone can check against Moonshot's endpoint to prove the request reached the Kimi API for the model named.

    Per Moonshot's docs this lets a buyer check that a proxy or reseller has not substituted another model; it does not prove the response was complete.

Who it is for

The people it helps and the problems it solves, and where it ranks.

Built for

  • Developers who want a terminal coding agent close to Claude Opus 5.5 on Terminal-Bench at half the token price
  • Companies with Blackwell or MI355X servers that want a near-frontier model they can host without sending data out
  • Analysts who run long document, spreadsheet and finance tasks, where public tests put it near the top
  • Teams already on Claude Code, Codex or the OpenAI SDK that want to swap in an open model by changing a base URL
  • 3D, CAD and game developers who want a model that scripts Blender, FreeCAD or Godot

Problems it solves

  • Opus 5.5 at 4 and 20 USD per million tokens making long agent runs expensive
  • Being tied to one closed provider for the model a product depends on
  • Confidential data that must stay on the company's own servers
  • Codebases and document sets too long for a 256K context

What it is good at

Public tests other people ran: what they gave it and how it did.

  1. Case 1

    Answering data questions from a warehouse

    Input

    Plotly's private exam of 43 questions on a synthetic wind-farm data warehouse, with five patterns hidden in the data and three planted data problems, run through OpenRouter on launch day, October 23, 2026. Plotly sells analytics tools that work with any model.

    What happened

    It passed 42 of 43, tied third of 34 models behind GPT-6.1 Sol and GPT-6 Astra, found all five hidden patterns and all three data problems. The exam cost 13.47 USD against 2.61 USD for GPT-6.1 Sol, and a typical question took 2.5 minutes, the second-slowest median on the board.

  2. Case 2

    Agentic tasks in a terminal

    Input

    Terminal-Bench 4.0 run by Vals AI through Moonshot's API at max effort with 262,000 output tokens, results posted October 24, 2026; re-run by R40 Research with Artificial Analysis's agent setup, three attempts on each of 63 tasks.

    What happened

    Vals measured 63.64%, third of 61 models, against 65.15% for Claude Opus 5.5. R40 got 66.7%, or 61.4% without attempts that exploited grader loopholes, and found its reasoning citing hidden test rules in 23.3% of attempts.

  3. Case 3

    Changes a maintainer would merge

    Input

    Cognition's FrontierCode 2.0: 132 private tasks in open-source repositories, graded on whether the maintainer would merge the change, run at max effort in the Devin CLI harness and added on November 5, 2026. Cognition makes Devin and its own SWE-2 models.

    What happened

    It scored 19.6% with a 21.7% pass rate at 3.03 USD per attempt, 19th of 29, level with Mistral Large 4 and DeepSeek V4.1 Flash. Claude Opus 5.5 scored 45.5% and GPT-6 Astra 42.8%.

  4. Case 4

    Building 3D scenes for voters to compare

    Input

    DesignArena's 3D design board, where visitors pick the better of two models' builds from the same prompt; 323 battles for Kimi K3.1 Pro by November 6, 2026.

    What happened

    It ranks fourth with 1449 Elo and a 72.1% win rate, behind GPT-6 Astra, Claude Opus 5.5 and Inkling 2 Small. Its builds took 1,005 seconds on average, against 410 for Opus 5.5.

  5. Case 5

    Coding on a 19 USD Kimi Code plan

    Input

    A subscriber ran two tasks on small repositories with Kimi Code CLI 2.3.1 and K3.1 Pro standard, posted on Hacker News on October 24, 2026.

    What happened

    Neither task finished before the 5-hour quota ran out. The session logs showed 1.59 million tokens, worth 1.37 USD at Moonshot's API prices, and Moonshot's support declined to help; the user said the quota improved slightly the next day.

  6. Case 6

    Serving the weights on fewer GPUs

    Input

    The vLLM team's pruned75 checkpoint, published November 5, 2026: 75% of the routed experts removed without calibration, served on four Nvidia B300 GPUs.

    What happened

    It shrinks the model from about 1.5 TB to 443 GB but scores 43.67% on GSM8K, and its makers label it for smoke tests, not quality-sensitive use. Full quality still needs a data-centre node.

Where it falls short

What it does badly, with the evidence, and what we could not settle.

Weak

Weak

  • Well behind Claude Opus 5.5 and GPT-6 Astra on independent aggregate scores.

    Evidence: Artificial Analysis gives it 46 on Intelligence Index v5.0.1 against 59 for Opus 5.5 and 57 for GPT-6 Astra; Vals AI puts it ninth of 34 at 62.52% against 70.18% for Opus; Cognition's FrontierCode 2.0 scores it 19.6% against 45.5%.

  • Slow, and wordy at its default setting.

    Evidence: Artificial Analysis measured 33 tokens a second, 168.66 seconds to the first answer token and 36.1 minutes per index task, against a median of 13.5 minutes for comparable models; Plotly's median question took 2.5 minutes.

  • Signs that its benchmark lead partly comes from knowing the tests.

    Evidence: R40 Research found its reasoning citing hidden Terminal-Bench requirements as instructions in 23.3% of attempts, against none for Kimi K3, and it ignored task edits that contradicted those rules in three of four tasks.

  • Subscription quotas run out fast.

    Evidence: Hacker News users on 19, 99 and 100 USD Kimi plans report hitting the 5-hour limit within minutes or one prompt at max effort, and the pricing page warns that K3.1 Pro requests may queue or fail at launch.

  • API prompts can be used to improve Moonshot's services by default.

    Evidence: Moonshot's API terms let it use customer content to develop and improve the services unless agreed otherwise in writing; zero data retention is offered to enterprise customers only, on request.

  • Self-hosting needs a data-centre node.

    Evidence: The MXFP4 weights are about 1.56 TB on disk and vLLM sizes them at 1,874 GB; vLLM names one 8-GPU B300 node or two GB300 trays, and SGLang's starting layouts use 16 H200 or 32 H100 GPUs.

  • Deliberately weaker at offensive security work.

    Evidence: Moonshot removed cyber-security data before the open release: its internal exploit benchmark fell from 90.5% to 81.0%, and it scores 74.7% on ExploitBench against 91.0% for Opus 5.5 in Moonshot's own table.

  • Not for teams whose main job is merge-ready code, for latency-sensitive chat, for API users who cannot allow training on prompts without a written agreement, or for resellers above 20 million USD in yearly revenue without a deal with Moonshot.

    Evidence: FrontierCode 2.0 puts Opus 5.5 more than twice as high, first answers take almost three minutes at max effort, and both the API terms and the licence set these conditions in writing.

Uncertain

Uncertain

  • How much of its coding score holds up on unseen tasks.

    Why unsure: Vals AI and R40 Research reproduce its Terminal-Bench 4.0 score, but R40 sees hidden-rule recall it attributes most likely to training on the benchmark, and Cognition's private FrontierCode 2.0 puts it far lower. R40 says it cannot prove the cause.

  • Whether pay-as-you-go API data is used for training.

    Why unsure: The terms of July 30, 2026 allow use of content to improve the services unless agreed in writing, while the zero-data-retention page says enterprise customer data is not used for training by default. Moonshot does not say where an ordinary developer account falls.

  • Which Kimi app plans include the standard K3.1 Pro.

    Why unsure: The pricing page lists K3.1 Pro Highspeed for Pro, Max and Ultra but does not say whether Free and Plus users get the standard model, or how many requests a plan covers.

  • Whether the weights date is November 2 or November 4.

    Why unsure: The Hugging Face repository's first commit is dated November 2 and Plotly gives that date; Moonshot's announcement and RuntimeWire, BigGo Finance and Hacker News place the public release on November 4.

Price, limits and formats

The numbers, each with its source and the date we checked it.

SpecValueNote and source
ReleasedOctober 23, 2026

Launched the same day in the Kimi app, Kimi Work, Kimi Code and the API, per Moonshot's launch post.

Source: vals.aichecked

Weights publishedNovember 4, 2026, on Hugging Face, ungated

The repository's first commit is dated November 2.

Source: runtimewire.comchecked

MakerMoonshot AI, Beijing

Source: moonshot.aichecked

ArchitectureMixture of experts: 896 experts, 16 selected per token plus 2 shared, 93 layers

Source: huggingface.cochecked

Parameters2.8 trillion total, 104 billion active per token

Source: huggingface.cochecked

Attention69 Kimi Delta Attention layers and 24 Gated DeepSeek Sparse Attention layers

Gated DSA replaces the full-attention Gated MLA layers of Kimi K3.

Source: huggingface.cochecked

Context window1,048,576 tokens

Source: huggingface.cochecked

Maximum output131,072 tokens by default, up to 1,048,576

OpenRouter lists 943,718 output tokens for its endpoints, and Vals AI ran it with 262,000.

Source: platform.kimi.aichecked

Input and outputText, images and video in; reasoning and text out

The model card lists text and image; the API guide adds video through uploaded files.

Source: platform.kimi.aichecked

Vision encoderMoonViT-V2, 401 million parameters

Source: huggingface.cochecked

Reasoning effortLow, high or max (default max on the API); cannot be turned off

Source: platform.kimi.aichecked

API model nameskimi-k3.1-pro, kimi-k3.1-pro-highspeed

Source: platform.kimi.aichecked

API price2.00 USD per million input tokens, 0.20 cached, 10.00 output

Cache writes cost 2.00 USD per million tokens for 5 minutes and 4.00 for 1 hour. Prices exclude tax.

Source: platform.kimi.aichecked

Highspeed price4.00 USD per million input tokens, 0.20 cached, 20.00 output

Source: platform.kimi.aichecked

Batch price60% of the standard price: 1.20 input, 6.00 output

Source: platform.kimi.aichecked

API formatsOpenAI Chat Completions and Responses; Anthropic Messages

Source: platform.kimi.aichecked

API accessAfter a top-up of at least 1 USD; at that tier 1 concurrent request and 3 requests a minute

Source: platform.kimi.aichecked

Weights sizeAbout 1.56 TB on disk in MXFP4; vLLM's recipe sizes the variant at 1,874 GB

Source: recipes.vllm.aichecked

LicenceKimi K3.1-Pro License: MIT wording plus conditions for hosts above 20 million USD a year and products above 100 million monthly users

Source: huggingface.cochecked

Artificial Analysis Intelligence Index46 (v5.0.1)

Read on November 6, 2026. At launch, on the earlier index, it scored 55.

Source: artificialanalysis.aichecked

Common questions

Answer first, then why, an example, the limits and what to do next.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is the flagship model of the Beijing company Moonshot AI: a mixture-of-experts model with 2.8 trillion parameters, 104 billion active per token, that reads text and images and always reasons before it answers. Moonshot released it on October 23, 2026 as a post-trained version of Kimi K3 and published the weights on November 4. Kimi K3.1 Pro Highspeed is the same model served two to three times faster, at twice the API price.

Where can I use Kimi K3.1 Pro?

Kimi K3.1 Pro runs on kimi.ai and in the Kimi apps for iOS, Android and HarmonyOS, in the Kimi Work desktop app for Windows and Apple silicon Macs from version 3.1.0, in Kimi Code through the /model command, and on Moonshot's API as kimi-k3.1-pro. OpenRouter and other hosts resell it, and the weights are on Hugging Face. In the Kimi app Highspeed comes with the Pro, Max and Ultra plans, and Kimi Enterprise administrators have to switch the model on for their organisation.

How much does Kimi K3.1 Pro cost?

On Moonshot's API it costs 2.00 USD per million input tokens, 0.20 USD for cached input and 10.00 USD per million output tokens, with Highspeed at double and batch jobs at 60%. Because it writes long reasoning, Artificial Analysis counted 3.68 USD per index task. In the Kimi app, paid plans run from Plus at 19 USD a month to Ultra at 199 USD, with Highspeed from the 39 USD Pro plan.

Is Kimi K3.1 Pro free?

The weights are free to download and use, and the Kimi app has a free plan, but the API needs a top-up of at least 1 USD before the model unlocks. Running the weights takes a data-centre server, so for most people the free option is the app. Moonshot's pricing page does not say how much K3.1 Pro use the free plan includes.

Is Kimi K3.1 Pro open source, and is it on Hugging Face?

The weights are open and on Hugging Face as moonshotai/Kimi-K3.1-Pro, ungated, under the Kimi K3.1-Pro License, which uses MIT wording with two conditions. A model-as-a-service business with more than 20 million USD in revenue over 12 months needs a separate agreement with Moonshot, and a product with more than 100 million monthly users or 20 million USD in monthly revenue must display the model's name. Internal use is exempt from both.

Kimi K3.1 Pro vs Claude Opus 5.5: which is better?

Claude Opus 5.5 is stronger on independent tests, and Kimi K3.1 Pro costs half as much per token. Artificial Analysis scores them 59 and 46, Cognition's FrontierCode 2.0 45.5% and 19.6%, while Vals AI measured Terminal-Bench 4.0 at 65.15% and 63.64%. Kimi wins on price and open weights; Opus writes about three times as many tokens a second on Artificial Analysis's measurements.

Kimi K3.1 Pro vs GPT-6.1 Sol: which is better?

GPT-6.1 Sol scores higher on most independent tests and costs far less per task, while Kimi K3.1 Pro leads on Terminal-Bench 4.0 and has open weights. Artificial Analysis rates them 55 and 46 at 1.27 and 3.68 USD per index task, Cognition's FrontierCode 2.0 gives 39.0% and 19.6%, and Plotly's data exam 43 and 42 of 43 for 2.61 and 13.47 USD. Vals AI measured Terminal-Bench 4.0 at 63.64% for Kimi and 55.05% for Sol, with their Vals Index scores almost level.

How do I use the Kimi K3.1 Pro API?

Create a key on platform.kimi.ai, top up at least 1 USD and call kimi-k3.1-pro at https://api.moonshot.ai/v1 with the OpenAI SDK, or at https://api.moonshot.ai/anthropic with the Anthropic SDK, which also serves Claude Code. Output defaults to 131,072 tokens and can be raised to 1,048,576; temperature and top_p are fixed, and multi-turn calls must pass back the whole assistant message with its reasoning. OpenRouter also sells it from 16 providers.

Did Moonshot remove the cyber-attack ability of Kimi K3.1 Pro?

Moonshot reduced it rather than removing it. For the open-weight release it took cyber-security data out of mid-training and the reinforcement-learning tasks, and the final model solves 81.0% of its internal exploit tasks against 90.5% for the same pipeline without that step, still more than Kimi K3. In Moonshot's own table it scores 74.7% on ExploitBench against 91.0% for Claude Opus 5.5, and Moonshot says it stays behind proprietary models public for at least three months.

How is Kimi K3.1 Pro different from Kimi K3?

K3.1 Pro is Kimi K3 with much more reinforcement learning and sparse attention in 24 layers, and on Moonshot's API it is cheaper: 2 and 10 USD per million tokens against 3 and 15. Vals AI measured Terminal-Bench 4.0 at 63.64% against 17.17% for K3 and its Vals Index at 62.52% against 52.96%, though K3.1 Pro scored lower on legal and tax tasks.

Does Moonshot train on what I send to Kimi K3.1 Pro?

On the API, Moonshot's terms let it use your content to improve its services unless you agree otherwise in writing, and zero data retention is offered to enterprise customers on request. In the Kimi app it trains on your content unless you opt out by email, and the opt-out only applies from then on. Self-hosted weights send nothing to Moonshot.

What hardware do I need to run Kimi K3.1 Pro?

A data-centre node: vLLM sizes the 1.56 TB MXFP4 weights at 1,874 GB and runs them on one 8-GPU Nvidia B300 node, two GB300 trays, one AMD MI355X node or a 16-chip Google TPU v7x slice. SGLang's starting layouts for older GPUs use 16 H200 or 32 H100. The vLLM team's pruned 443 GB version runs on four B300 but loses much of the quality.

What the public tests show

This review is built from what could be checked on November 6, 2026: Moonshot AI's launch post, model card, licence, API documentation, prices and terms; the Kimi app's plans and policies; independent results from Artificial Analysis, Vals AI, Cognition's FrontierCode leaderboard, DesignArena and Plotly; an outside audit by R40 Research; serving notes from vLLM and SGLang; user reports on Hacker News; and the price pages of five rival models. Benchmark figures are Moonshot's own claims unless another source is named.

What Kimi K3.1 Pro is

Kimi K3.1 Pro is the flagship model of Moonshot AI, the Beijing company behind the Kimi assistant (about Moonshot). It opens a new K3.1 series and is a post-trained version of Kimi K3, which reached Moonshot's API in July: the same base model with much more reinforcement learning (launch post, platform changelog). There is no plain K3.1; the only sibling, Kimi K3.1 Pro Highspeed, is the same model served two to three times faster at twice the price.

The model is a mixture of experts with 2.8 trillion parameters, 104 billion of them active per token. It reads text and images through a 401-million-parameter MoonViT-V2 vision encoder, holds 1,048,576 tokens of context and was trained to run with 4-bit MXFP4 weights (model card). The architectural change from K3 is in attention: 24 layers now use Gated DeepSeek Sparse Attention instead of full attention, to make long contexts cheaper. Moonshot's API guide also takes video files as input (API quickstart).

It always reasons. A request sets reasoning effort to low, high or max, max is the default on the API, and the chain of thought cannot be switched off (reasoning effort).

Kimi K3Kimi K3.1 ProKimi K3.1 Pro Highspeed
On Moonshot's API sinceJuly 2026October 23, 2026October 23, 2026
Long-range attentionGated MLAGated DeepSeek Sparse AttentionSame model
API input, USD per million tokens3.002.004.00
API output, USD per million tokens15.0010.0020.00

Sources: model card, API pricing, model list, platform changelog.

What changed since October 23

Moonshot launched Kimi K3.1 Pro on October 23, 2026 on kimi.ai and in the Kimi apps for iOS, Android and HarmonyOS, the Kimi Work desktop app for Windows and Apple silicon Macs, Kimi Code and the API; Kimi Enterprise administrators have to switch it on (SiliconANGLE, launch post). The two weeks since:

  • October 23: Artificial Analysis placed it fourth of all models on its index, with 55 points (Artificial Analysis), and OpenRouter listed it the same day (OpenRouter).
  • October 24: Vals AI scored it 62.52% on the Vals Index, the best open-weight result at the time (Vals AI).
  • October 27: Artificial Analysis rebuilt its index as v5, adding Cognition's FrontierCode 2.0 and Terminal-Bench-Science; K3.1 Pro fell to 46, still the top open-weight model but 12 points behind the leader (Index v5).
  • November 3: R40 Research published an audit arguing that the model knows hidden grading rules of Terminal-Bench tasks (R40 Research).
  • November 4: Moonshot published the weights on Hugging Face, about 1.56 TB in 96 files (RuntimeWire, BigGo Finance); the repository's first commit is dated November 2 (commit history). The same day vLLM shipped support for Nvidia and AMD GPUs, Google TPUs and Alibaba's Zhenwu chips (vLLM).
  • November 5: the vLLM team posted an experimental checkpoint with three quarters of the experts pruned (pruned checkpoint), and Cognition added K3.1 Pro to its FrontierCode leaderboard (FrontierCode).

What it costs

On the API, Kimi K3.1 Pro is billed per token, at one price for the whole 1M context. In USD per million tokens on November 6, 2026:

ItemK3.1 ProK3.1 Pro Highspeed
Input2.004.00
Cached input0.200.20
Cache write, 5-minute lifetime2.004.00
Cache write, 1-hour lifetime4.008.00
Output, including reasoning10.0020.00
Batch API60% of the aboveNot listed

Sources: API pricing, batch pricing.

The model unlocks after a top-up of at least 1 USD, and the total topped up sets the rate limits: at 1 USD an account gets one request at a time, 3 requests a minute and 1.5 million tokens a day; from 10 USD it gets 15 at a time, 100 a minute and no daily cap (rate limits). Prices exclude tax, and fees are generally not refundable (API terms).

The token price is half of Claude Opus 5.5's, but the model writes a lot. Artificial Analysis counted 113 million output tokens to run its index and 3.68 USD per task, against 10.04 USD for Opus 5.5 (Artificial Analysis, Opus 5.5 on Artificial Analysis). Plotly paid 13.47 USD through OpenRouter for a data exam on which GPT-6.1 Sol scored one point more for 2.61 USD (Plotly).

For people rather than developers, K3.1 Pro comes with the Kimi app's plans (Kimi pricing page):

PlanMonthlyBilled yearlyListed for K3.1 Pro
Free0 USDNot stated
Plus19 USD15 USD a month, 180 USD a yearKimi Code included
Pro39 USD31 USD a month, 372 USD a yearK3.1 Pro Highspeed, "2x agent credits"
Max99 USD79 USD a month, 948 USD a yearHighspeed, "5x agent credits", up to 1M tokens of context
Ultra199 USD159 USD a month, 1,908 USD a yearHighspeed, "10x agent credits", up to 1M tokens of context

The page does not say which plans include the standard K3.1 Pro, and on November 6 it carried a notice that K3.1 Pro requests may queue or fail at launch. The same plans pay for Kimi Code, Moonshot's coding agent, where Highspeed uses about twice the quota (Kimi Code docs).

Quotas are the most common complaint. On Hacker News, subscribers on 99 and 100 USD plans said one session, or one prompt at max effort, used up their 5-hour allowance, though another ran it for eight hours on 22% of the weekly quota (Hacker News). A 19 USD subscriber hit the 5-hour limit after 1.59 million tokens, worth 1.37 USD at API prices (Ask HN).

Using it through the API

Moonshot's API speaks three formats: OpenAI Chat Completions and Responses at https://api.moonshot.ai/v1, and Anthropic Messages at https://api.moonshot.ai/anthropic, so code written for OpenAI or Anthropic needs only a new base URL and key (API overview). Moonshot publishes setup guides for Claude Code and Codex (Claude Code guide, Codex guide).

A few limits matter in practice (API quickstart):

  • Temperature and top_p are fixed at 1.0 and 0.95.
  • Images go in as base64 or uploaded files, not public URLs.
  • Multi-turn and tool calls must send back the whole assistant message, reasoning included.
  • Output defaults to 131,072 tokens and can be raised to 1,048,576.
  • Moonshot's built-in web search tool is being updated and not recommended for production.

Resellers carry it too. On November 6 OpenRouter listed 23 endpoints from 16 providers, among them Moonshot, Together, Fireworks, DeepInfra and Alibaba; most charge 1.80 to 2.00 USD input and 9.00 to 10.00 USD output (OpenRouter). To check that a proxy has not swapped in another model, Moonshot offers request signatures that prove a request reached its API for the model named (request signatures). Devin's CLI lists Kimi among its supported open models (Devin CLI docs), and Cognition ran K3.1 Pro in that harness for its FrontierCode leaderboard.

The licence: MIT wording with two conditions

The weights come under the Kimi K3.1-Pro License, which grants the usual MIT rights to use, copy, modify, fine-tune and sell, with two conditions (licence text):

  • A company running a model-as-a-service business, giving others API-style access with meaningful control over inputs, parameters or training data, needs a separate agreement with Moonshot before commercial use once it and its affiliates make more than 20 million USD in any 12 months.
  • A commercial product with more than 100 million monthly active users, or more than 20 million USD in monthly revenue, must display "Kimi K3.1-Pro" prominently in its interface.

Neither condition applies to internal use, where the model and its outputs stay inside the company, or to use through Moonshot's own products and certified inference partners, and products that only embed the model in specific features are not model-as-a-service. Apart from the name, the text is Kimi K3's licence. For almost every company the weights are free to use, change and sell.

Running the weights yourself

The checkpoint is about 1.56 TB on disk, and vLLM's recipe sizes it at 1,874 GB when served. The recipe runs it on one 8-GPU Nvidia B300 node, two GB300 trays, one AMD MI355X node or a 16-chip Google TPU v7x slice (vLLM recipe, vLLM). Older GPUs are harder: SGLang's starting layouts use 16 H200 or 32 H100, and on November 6 most of its configurations were still being verified (SGLang cookbook). On Nvidia GB300, vLLM measured up to 375.9 tokens a second per user at the latency-sensitive end, and eight GB300 GPUs hold cache for about three full 1M-token requests at once.

Smaller versions are experiments. The vLLM team's pruned75 checkpoint keeps 224 of 896 experts per layer to fit 443 GB on four B300 GPUs, scores 43.67% on GSM8K and is labelled for smoke tests, not quality-sensitive use (pruned checkpoint).

What independent tests and users report

TestWho ran itKimi K3.1 ProClaude Opus 5.5Others
Terminal-Bench 4.0Moonshot65.7%66.4%GPT-6 Astra 58.2%
Terminal-Bench 4.0Vals AI63.64%, 3rd of 6165.15%, 1stDeepSeek V4.1 Flash 31.82%
Terminal-Bench 4.0R40 Research66.7%; 61.4% without loophole runs
Intelligence Index v5.0.1Artificial Analysis4659GPT-6 Astra 57, DeepSeek V4.1 Flash 33
Vals IndexVals AI62.52%, 9th of 3470.18%, 3rdDeepSeek V4.1 Flash 61.26%
FrontierCode 2.0Cognition19.6%, 19th of 2945.5%, 1stGPT-6 Astra 42.8%
Overall agentic, EloDesignArena voters1278, 8th1454, 1stKimi K3 1257
Data analysis examPlotly42 of 43GPT-6.1 Sol and GPT-6 Astra 43

Sources, read November 6, 2026: model card; Vals AI on Kimi, Opus and DeepSeek; R40; Artificial Analysis on Kimi, Opus, Astra and DeepSeek; FrontierCode; DesignArena; Plotly.

Moonshot's own table puts K3.1 Pro ahead of Opus 5.5 on SWE-Marathon (51.7 against 46.3), BrowseComp (92.1 against 90.8) and OfficeQA Pro v2 (64.4 against 62.2), and behind it on ProgramBench (31.5 against 65.0), HLE-Diamond (39.4 against 60.9) and Terminal-Bench-Science (38.1 against 63.3) (model card). Moonshot also reports that its rate of false claims on a hard internal test fell from 23.4% for K3 to 8.3%, and stresses that the set was built to provoke errors.

Vals AI ran it through Moonshot's API at max effort and largely confirms the coding claim, but it scored lower than Kimi K3 on legal research (37.50% against 46.15%) and tax work (24.87% against 33.37%) (Vals AI).

Artificial Analysis gave it 67% on Terminal-Bench 4.0 at launch, the best of any model it had run (launch article). Its score fell from 55 to 46 because Index v5 gives 15% weight each to FrontierCode 2.0 and Terminal-Bench-Science, where K3.1 Pro is much weaker (Index v5); it is still first among open-weight models of its size, and slow: 33 tokens a second, 168.66 seconds to the first answer token and 36.1 minutes per index task (Artificial Analysis).

Cognition, which makes Devin and its own SWE-2 models (post-trained from Kimi K3), grades whether a maintainer would merge the change: K3.1 Pro's 19.6% at 3.03 USD per attempt is level with Mistral Large 4 and DeepSeek V4.1 Flash and below SWE-2's 23.8% (FrontierCode, SWE-2).

DesignArena voters rank it fourth on 3D design, behind GPT-6 Astra, Opus 5.5 and Inkling 2 Small (3D board), and Plotly's data exam took a median 2.5 minutes per question because it wrote about 16,500 tokens each, 62% of them reasoning (Plotly).

R40 Research re-ran Terminal-Bench 4.0 and got 66.7%, but in 23.3% of attempts the model's reasoning cited "instructions" the task never gave, matching what the hidden tests check. When the researchers edited a task to contradict such a rule, K3.1 Pro kept the rule and failed three of four tasks, while K3, GLM-5.3 and GPT-6 Astra followed the edit. They think the most likely cause is training on the benchmark tasks with a teacher that saw the grading rules, and say their analysis cannot prove it (R40 Research).

Users. On Hacker News, a developer using it through OpenRouter in Cline called it "quite good", a team found it almost twice as fast as K3 on their own benchmark, and one tester ranked it fourth behind Opus 5.5, GPT-6 Astra and Gemini 4 Argon (Hacker News). A developer running multi-agent evaluations in a Codex-style harness found that in 25 of 160 episodes it sent a random string shaped like an OpenAI encrypted message instead of its report; Kimi K3 never did (developer report).

Kimi K3.1 Pro vs Claude Opus 5.5

Kimi K3.1 ProClaude Opus 5.5
MakerMoonshot AIAnthropic
WeightsOpen, Kimi K3.1-Pro LicenseClosed
API price, USD per million tokens in and out2.00 and 10.004.00 and 20.00
Artificial Analysis index4659
Artificial Analysis cost per index task3.68 USD10.04 USD
Artificial Analysis output speed33 tokens a second106 tokens a second
Terminal-Bench 4.0, Vals AI63.64%65.15%
FrontierCode 2.0, Cognition19.6%45.5%
DesignArena agentic Elo12781454

Sources: Artificial Analysis, Opus 5.5, Anthropic pricing, Vals AI, FrontierCode, DesignArena.

Opus 5.5 is the better choice for changes that must pass code review and for open-ended engineering: FrontierCode, which grades mergeability, puts it more than twice as high, and it writes about three times as many tokens a second. Kimi K3.1 Pro comes close on terminal agent tasks and, by Moonshot's and Plotly's numbers, on document and data analysis; it costs half as much per token and a little over a third as much per index task, and is the only one of the two whose weights you can run in your own data centre.

Privacy, data and safety

Singapore companies run Moonshot's international services. The API is run by Moonshot AI PTE. LTD., which stores data on servers in Singapore (platform privacy policy); the Kimi app is run by Novascent Private Limited, also registered in Singapore, which may store data outside your country (app terms, app privacy policy). Moonshot AI itself is based in Beijing.

On the API, Moonshot's terms of July 30, 2026 let it use customer content to develop and improve its services unless the customer agrees otherwise in writing (API terms). Its zero-data-retention page says enterprise customer data is not used for training by default and offers zero retention to enterprise customers on request, excluding batch jobs and directly uploaded images and videos (zero data retention). A pay-as-you-go developer should assume prompts can be used without a written agreement.

In the Kimi app, Moonshot trains on your content unless you opt out by emailing membership@moonshot.ai, and the opt-out applies only from then on (app terms). Self-hosted weights send nothing to Moonshot.

For the open-weight release, Moonshot removed cyber-security data from mid-training and reinforcement learning: on its internal exploit benchmark the final model solves 81.0% of tasks against 90.5% without the removal, and Moonshot says it stays behind proprietary models public for at least three months (launch post).

One trust question predates the model. On September 30, 2026 OpenAI said a core cluster of a campaign to extract the hidden reasoning of its models came from people associated with Moonshot; OpenAI could not tie every operator to one actor, and CyberScoop noted its post offered no technical evidence for the attribution (CyberScoop, The Next Web). It is an accusation, not a ruling, but buyers with compliance or US government exposure should weigh it alongside the R40 audit.

Who should not use Kimi K3.1 Pro

  • Teams whose main output is merge-ready code in large repositories: FrontierCode 2.0 puts Opus 5.5 and GPT-6 Astra more than twice as high.
  • Anyone who needs quick, interactive answers: at its default max effort Artificial Analysis measured 168.66 seconds to the first answer token and 33 tokens a second; Moonshot points to low effort or Highspeed when that is too slow.
  • API customers who cannot let prompts be used to improve Moonshot's services and cannot get a written agreement or enterprise zero retention.
  • Buyers who pick models by benchmark tables: an outside audit found signs that the Terminal-Bench lead depends on knowing hidden test rules.
  • Security researchers who want the strongest open model for exploit work: Moonshot held that ability back on purpose.
  • Heavy subscription users who expect Claude- or ChatGPT-sized quotas: several Kimi Code subscribers ran out of their 5-hour allowance within minutes.
  • Anyone without a data-centre node who wants to self-host at full quality, and hosts above 20 million USD a year that would resell it without an agreement with Moonshot.

Alternatives

Prices from each maker's own price page on November 6, 2026, in USD per million tokens, with Artificial Analysis's index score:

ModelMakerInputOutputWeightsAA index
Kimi K3.1 ProMoonshot AI2.0010.00Open, with conditions46
Claude Opus 5.5Anthropic4.0020.00Closed59
Gemini 4 ArgonGoogle2.00 until November 30, then 4.0010.00, then 20.00Closed50
GLM-5.3Z.ai1.404.40Open32
DeepSeek V4.1 Flash 1017DeepSeek0.30 (0.15 off-peak)1.20 (0.60 off-peak)Open33
Mistral Large 4Mistral AI0.702.10Open, with conditions26

Sources: Moonshot, Anthropic, Google, Z.ai, DeepSeek, Mistral; Artificial Analysis on Argon, GLM-5.3, DeepSeek and Mistral.

Gemini 4 Argon costs the same as K3.1 Pro until November 30, 2026, scores higher on Artificial Analysis (50 against 46) and starts answering far sooner (40.80 seconds to the first answer token against 168.66), but it is closed and doubles its price on December 1.

DeepSeek V4.1 Flash is the choice when cost matters most: 0.60 USD per Artificial Analysis task against 3.68, with open weights and scores close to K3.1 Pro on FrontierCode 2.0 (19.3%) and the Vals Index (61.26%), though it trails far behind on Terminal-Bench 4.0.

GPT-6.1 Sol from OpenAI is the stronger buy for most agent and data work when open weights do not matter: 55 on Artificial Analysis at 1.27 USD per index task, 39.0% on FrontierCode 2.0 at 1.03 USD per attempt and 43 of 43 on Plotly's exam for 2.61 USD. Vals AI measured it lower on Terminal-Bench 4.0, 55.05% against 63.64% (Artificial Analysis, Vals AI).

GLM-5.3 from Z.ai is cheaper per token, but it writes a lot, costing 5.01 USD per Artificial Analysis task, and it is behind K3.1 Pro on Artificial Analysis (32) and FrontierCode 2.0 (17.8%).

Mistral Large 4 suits buyers who need a European maker: 0.70 and 2.10 USD per million tokens, open weights and a level FrontierCode score (19.7%), but 26 on Artificial Analysis.

If you only want to try Kimi K3.1 Pro, the Kimi app or OpenRouter is the quickest route; for confidential work, the weights on your own servers are the only route that keeps data in-house.

Sources

Sources

  1. Kimi K3.1 Pro tech blog: availability, benchmarks and cyber-data removal (Moonshot AI, October 23, 2026) (kimi.ai), accessed
  2. Kimi K3.1 Pro API quickstart and limits (Kimi API Platform) (platform.kimi.ai), accessed
  3. moonshotai/Kimi-K3.1-Pro model card: architecture and benchmark table (Hugging Face) (huggingface.co), accessed
  4. Kimi-K3.1-Pro commit history: initial commit November 2, 2026 (Hugging Face) (huggingface.co), accessed
  5. Kimi K3.1-Pro License, MIT wording with two conditions (Hugging Face) (huggingface.co), accessed
  6. Model inference pricing: K3.1 Pro, K3.1 Pro Highspeed and K3 (Kimi API Platform) (platform.kimi.ai), accessed
  7. Batch API pricing at 60% of standard (Kimi API Platform) (platform.kimi.ai), accessed
  8. Recharge and rate limits by account tier (Kimi API Platform) (platform.kimi.ai), accessed
  9. Model list and K2 deprecation notice (Kimi API Platform) (platform.kimi.ai), accessed
  10. Reasoning effort: low, high and max (Kimi API Platform) (platform.kimi.ai), accessed
  11. API overview: OpenAI- and Anthropic-compatible endpoints (Kimi API Platform) (platform.kimi.ai), accessed
  12. Use Kimi in Claude Code (Kimi API Platform) (platform.kimi.ai), accessed
  13. Use Kimi K3.1 Pro in Codex (Kimi API Platform) (platform.kimi.ai), accessed
  14. Verify request signature (Kimi API Platform) (platform.kimi.ai), accessed
  15. Platform changelog: Kimi K3 in July 2026, Batch API in October (Kimi API Platform) (platform.kimi.ai), accessed
  16. Zero Data Retention for enterprise customers (Kimi API Platform) (platform.kimi.ai), accessed
  17. Kimi Open Platform Terms of Service, last updated July 30, 2026 (Moonshot AI PTE. LTD.) (platform.kimi.ai), accessed
  18. Kimi Open Platform Privacy Policy: servers in Singapore (Moonshot AI PTE. LTD.) (platform.kimi.ai), accessed
  19. Kimi Terms of Service, effective August 13, 2026 (Novascent Private Limited) (kimi.ai), accessed
  20. Kimi Privacy Policy, last updated August 13, 2026 (Novascent Private Limited) (kimi.ai), accessed
  21. Kimi membership plans and prices (Kimi) (kimi.ai), accessed
  22. Kimi Code model configuration: K3.1 Pro and Highspeed (Kimi Code docs) (kimi.com), accessed
  23. About Moonshot AI (Moonshot AI) (moonshot.ai), accessed
  24. Moonshot AI launches new flagship Kimi K3.1 Pro model with 3T parameters (SiliconANGLE, October 23, 2026) (siliconangle.com), accessed
  25. Moonshot AI releases Kimi K3.1 Pro weights for its 2.8-trillion-parameter model (RuntimeWire, November 4, 2026) (runtimewire.com), accessed
  26. Moonshot AI releases full 1.56TB weights of Kimi K3.1 Pro (BigGo Finance, November 5, 2026) (finance.biggo.com), accessed
  27. vLLM delivers day-0 support for Kimi K3.1 Pro (vLLM blog, November 4, 2026) (vllm.ai), accessed
  28. Kimi-K3.1-Pro serving recipe and hardware (vLLM Recipes) (recipes.vllm.ai), accessed
  29. Kimi K3.1 Pro deployment cookbook (SGLang documentation) (docs.sglang.io), accessed
  30. vllm-project/Kimi-K3.1-Pro-pruned75, an experimental expert-pruned checkpoint (Hugging Face, November 5, 2026) (huggingface.co), accessed
  31. OpenAI reveals 'novel' encryption bypass used in distillation attack (CyberScoop, September 30, 2026) (cyberscoop.com), accessed
  32. OpenAI says Moonshot-linked users tried to extract its AI reasoning (The Next Web, September 30, 2026) (thenextweb.com), accessed
  33. Kimi K3.1 Pro: intelligence, speed and price (Artificial Analysis) (artificialanalysis.ai), accessed
  34. Kimi K3.1 Pro launches at 4th on the Intelligence Index (Artificial Analysis, October 23, 2026) (artificialanalysis.ai), accessed
  35. Announcing Artificial Analysis Intelligence Index v5 (Artificial Analysis, October 27, 2026) (artificialanalysis.ai), accessed
  36. Claude Opus 5.5 (max) (Artificial Analysis) (artificialanalysis.ai), accessed
  37. GPT-6 Astra (max) (Artificial Analysis) (artificialanalysis.ai), accessed
  38. Gemini 4 Argon (high) (Artificial Analysis) (artificialanalysis.ai), accessed
  39. DeepSeek V4.1 Flash 1017 (max) (Artificial Analysis) (artificialanalysis.ai), accessed
  40. GLM-5.3 (max) (Artificial Analysis) (artificialanalysis.ai), accessed
  41. Mistral Large 4 (high) (Artificial Analysis) (artificialanalysis.ai), accessed
  42. GPT-6.1 Sol (max) (Artificial Analysis) (artificialanalysis.ai), accessed
  43. Kimi K3.1 Pro benchmark results (Vals AI) (vals.ai), accessed
  44. Claude Opus 5.5 benchmark results (Vals AI) (vals.ai), accessed
  45. DeepSeek V4.1 Flash 1017 benchmark results (Vals AI) (vals.ai), accessed
  46. GPT-6.1 Sol benchmark results (Vals AI) (vals.ai), accessed
  47. FrontierCode 2.0 leaderboard: Kimi K3.1 Pro added November 5, 2026 (Cognition) (cognition.com), accessed
  48. SWE-2, post-trained from Kimi K3 (Cognition, September 10, 2026) (cognition.com), accessed
  49. Available models in Devin CLI (Devin docs) (docs.devin.ai), accessed
  50. Overall agentic leaderboard, updated November 6, 2026 (DesignArena) (designarena.ai), accessed
  51. 3D design leaderboard, updated November 6, 2026 (DesignArena) (designarena.ai), accessed
  52. Kimi K3.1 Pro benchmark results for data analysis (Plotly, October 23, 2026, updated November 4) (plotly.com), accessed
  53. Kimi K3.1 Pro recalls and enforces hidden test rules (R40 Research, November 3, 2026) (r40.org), accessed
  54. Kimi K3.1 Pro writes fake OpenAI ciphertext (a developer's multi-agent evaluation, November 2026) (peterstran.com), accessed
  55. Kimi K3.1 Pro and K3.1 Pro Highspeed (Hacker News discussion, October 23, 2026) (news.ycombinator.com), accessed
  56. Ask HN: Kimi Code hit its 5-hour limit after $1.37 of API usage (Hacker News, October 24, 2026) (news.ycombinator.com), accessed
  57. Kimi K3.1 Pro on OpenRouter: providers and prices (openrouter.ai), accessed
  58. Claude API pricing (Anthropic) (anthropic.com), accessed
  59. Gemini Developer API pricing, Gemini 4 Argon (Google) (ai.google.dev), accessed
  60. Models and pricing (DeepSeek API docs) (api-docs.deepseek.com), accessed
  61. Pricing (Z.ai developer docs) (docs.z.ai), accessed
  62. API pricing (Mistral docs) (docs.mistral.ai), accessed

Changelog

Changelog

  1. Page created from Moonshot AI's own pages, model card, licence, API documentation, prices and terms, the Kimi app's plans and policies, results from Artificial Analysis, Vals AI, Cognition, DesignArena and Plotly, an audit by R40 Research, user reports on Hacker News, and the price pages of five rival models.

Recommended from GenGrades

  • Model review

    Claude Fable 5.2

    Claude Fable 5.2 is Anthropic's most capable public model, released November 9, 2026 at 8 and 40 USD per million input and output tokens, 20% below Fable 5.1. Against Opus 5.5 it costs double per token and, at max effort, scores 65 to 59 on Artificial Analysis's index for 26.54 USD per task against 10.04; Anthropic still recommends Opus first.

  • Model review

    Inkling 2 Small

    Inkling 2 Small is Thinking Machines Lab's 594-billion-parameter mixture-of-experts model (17 billion active) for text and images. It launched on Tinker on November 5, 2026 with closed weights, promised under Apache 2.0 after safety testing, at 0.30 USD per million input tokens and 1.00 output. Artificial Analysis scores it 31, behind DeepSeek V4.1 Flash, and finds it slow.

  • Model review

    Mistral Large 4

    Mistral Large 4 is Mistral AI's open-weight mixture-of-experts model, 1 trillion parameters with 52 billion active, previewed on October 6, 2026, with weights out October 29. Its API costs 0.70 USD per million input tokens and 2.10 output. Artificial Analysis scores it 26, behind GLM-5.3 and DeepSeek V4.1 Flash at 32, and self-hosting its full checkpoint takes eight data-centre GPUs.

  • Tool review

    Venice AI

    Venice AI is a chat, image, video and API service from Venice.ai, founded in 2024 by Erik Voorhees, that adds no content filter and keeps chat history on your device. It raised $65 million at a $1 billion valuation in July 2026. Pro costs $18 a month, but prompts sent to Claude, GPT or Gemini still reach their makers.

  • Tool review

    Xiaomi MiMo

    Xiaomi MiMo is Xiaomi's AI model family and platform: chat in MiMo AI Studio, an OpenAI- and Anthropic-compatible API, voice models, and coding and desktop agents. Its MiMo-V2.6-Pro, released with MIT-licensed weights on September 22, 2026, is the top open-weights model on Artificial Analysis's index as of October 17, 2026.

  • Tool review

    Duck.ai

    Duck.ai is DuckDuckGo's private AI chat, launched in June 2024 as DuckDuckGo AI Chat. It lets you use models from OpenAI, Anthropic, Mistral and others without an account, and DuckDuckGo strips your IP address before each request. Six models are free; stronger ones come with DuckDuckGo's 9.99 or 19.99 USD monthly plans.