Kimi K3.1 Pro: what changed from K3 and what it really costs

Kimi K3.1 Pro already leads the open-weight models on Artificial Analysis’s index, ahead of the weights’ promised release by 4 November, though it trails Claude Opus 5.5 by 12 points on the new version of that index. Its tokens cost a third less than Kimi K3’s, but it writes so much that a finished task is not cheap: on Artificial Analysis’s tests it costs more per task than several models that score higher. For UK teams, the case for it rests on long agentic jobs run at lower effort settings, and on what the weights will let you do.

The answer in brief

• Kimi K3.1 Pro is Moonshot AI’s new flagship, released on 23 October 2026 alongside a faster K3.1 Pro Highspeed version. It is the first model in the K3.1 series; Moonshot’s model lists include nothing called simply Kimi K3.1.
• Moonshot describes it as Kimi K3 with substantially more post-training. The size (2.8 trillion parameters, 104 billion active per token) and the 1 million-token context window are unchanged.
• As of 31 October 2026 it scores 46 on version 5 of Artificial Analysis’s Intelligence Index, 12 points behind Claude Opus 5.5. Artificial Analysis still counts that as the top open-weight score, because Moonshot has promised to publish the weights. At launch, on the previous version, it ranked fourth and trailed Opus by three.
• The API costs $2 per million input tokens and $10 per million output tokens, list prices in US dollars, before tax. That is a third below K3 and the same base rates as Claude Sonnet 5.5.
• It writes a lot, so cheap tokens do not make cheap tasks. Artificial Analysis puts it at $3.68 per index task: about a third of Claude Opus 5.5 at its maximum setting, but more than Claude Opus 5.5 and GPT-6 Astra at their high settings, which both score higher.
• Its output is slow, at about 31 tokens a second on Artificial Analysis’s measurements. In our own tests, the Highspeed version ran two to three times as fast, at twice the standard input and output rates.
• The weights are not out yet. Moonshot says they will be released by 4 November 2026, and its launch materials do not name a licence.
• We would trial it for long coding-agent jobs and spreadsheet-heavy analysis at low or high effort, and compare the cost of finished tasks with your current model before moving any work.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a reasoning model from Moonshot AI, the Beijing company behind the Kimi assistant, and the first release in its K3.1 series. Moonshot’s launch post calls it its “most capable 3T-class open-weight model” and describes it as a post-trained version of Kimi K3, which launched on 16 July 2026.

It reads text, images and video. Moonshot says it always reasons before it answers: you set how hard it thinks with a reasoning effort of low, high or max, and thinking cannot be switched off.

Moonshot says K3.1 Pro Highspeed runs two to three times faster, and the Kimi Code documentation says it has the same capabilities. It costs twice as much per input and output token, although cache hits cost the same. You can use the models in:

• the Kimi website, with both models, and the Kimi apps for iOS, Android and HarmonyOS, for which Moonshot lists only K3.1 Pro
• the Kimi Work desktop app, version 3.1.0 or later, for Windows and Apple silicon Macs
• Kimi Code, Moonshot’s coding agent, through its /model command
• the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and OpenRouter as moonshotai/kimi-k3.1-pro

On OpenRouter, Highspeed is a separate endpoint that its standard routing does not use. For organisations on Kimi Business, the model starts switched off until an administrator enables it.

What changed from Kimi K3?

The training changed and the model’s size did not. Moonshot calls K3.1 Pro a post-trained version of K3, and its launch post on X gives the same 2.8 trillion parameters, 104 billion of them active, in a mixture-of-experts design; Artificial Analysis lists the same counts. Moonshot says the extra post-training ran large-scale reinforcement learning across hundreds of task domains.

The largest gains are in long, tool-using tasks. In Artificial Analysis’s launch analysis, K3.1 Pro scores 67% on Terminal-Bench 4.0, a test of coding agents working in a terminal, against 13% for K3; no other model Artificial Analysis has tested scores higher. On GDPval-AA, a test of real-world work tasks, it gains 312 Elo points to reach 1,846, third behind Claude Opus 5.5 and Meta’s Muse Wonder Full. Knowledge and science scores moved far less.

Source: Artificial Analysis launch analysis, 23 October 2026, on version 4.3 of its Intelligence Index, which version 5 replaced on 27 October. K3’s setting as Artificial Analysis labels it.

Moonshot’s own tests point the same way, though a vendor picks which benchmarks to publish. On its internal long-horizon coding test, Kimi Code Bench 3.0, the score rises from 55.95% to 61.2%. On its internal finance tests, K3.1 Pro gains 7.2 points on financial modelling (50.4 to 57.6) and 11.6 points on financial research (58.8 to 70.4), and scores 37 against K3’s 8.9 on a simulated business-management test.

Moonshot also worked on factual accuracy. Kimi Hallu Bench Hard is its set of prompts built to provoke factual errors. With tools enabled, the share of false claims fell from 23.4% with K3 to 8.3%. In the same run, GPT-6 Astra scored 5.5% and Claude Opus 5.5 15.0%. Moonshot calls this a stress test whose rates do not reflect everyday use. Its API quickstart reports 23.35% falling to 6.85% for a test it describes differently, built from user-flagged, de-identified conversations, so the two sets of figures are not directly comparable.

How does Kimi K3.1 Pro compare with Claude and GPT models?

On version 5 of Artificial Analysis’s Intelligence Index, launched on 27 October with harder and mostly private tests, K3.1 Pro scores 46. Claude Opus 5.5 leads the generally available models at 58, with GPT-6 Astra on 57 and GPT-6.1 Sol on 55 at their maximum settings; Google’s Gemini 4 Carbon scores 59 but is still rolling out to selected users. Artificial Analysis still counts K3.1 Pro as the top open-weight model, on the strength of Moonshot’s promise to publish the weights, but its gap to Claude Opus 5.5 has grown from 3 points to 12. Of models whose weights are already out, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32.

Source: Artificial Analysis leaderboard, Intelligence Index v5.0, checked 31 October 2026. Selected entries, with settings as Artificial Analysis labels them; it lists each effort setting separately and scores K3.1 Pro at a single setting it does not name. Cost per index task is Artificial Analysis’s weighted average for one task in its benchmark set. Gemini 4 Carbon is rolling out to selected users and is not yet publicly available; Artificial Analysis prices it at Google’s standard rates, without the 50% launch discount Google is offering.

The cost column is mixed. Against the other models at their maximum settings, K3.1 Pro’s $3.68 a task is about 37% of what Claude Opus 5.5 and Sonnet 5.5 cost and 59% of GPT-6 Astra’s, but nearly three times GPT-6.1 Sol’s $1.27, and all four score higher. Lower settings sharpen the contrast: at high, GPT-6 Astra scores 55 for $3.24 and Claude Opus 5.5 52 for $3.46, while GPT-6.1 Sol at medium matches K3.1 Pro’s 46 for $0.36. Its low token prices are undone by how much it writes: Artificial Analysis counted 110 million output tokens to run the index, against a median of 70 million for models in its price tier. Artificial Analysis also lists rivals at several settings and K3.1 Pro at one it does not name, so only the rivals’ cheaper settings can be picked out. Our explainers on what Claude Opus 5.5 changed and costs and on GPT-6 Astra for marketing cover those two models in more depth.

Moonshot’s own comparison, with K3.1 Pro at maximum effort, puts it at 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%). On BenchCAD, which scores how well a model rebuilds 3D objects as CAD code, it reaches 92.7% with tools, behind Opus 5.5 (96.2%) and Astra (95.9%). Moonshot claims the top score on OfficeQA Pro v2, a test of financial analysis across real documents, at 64.4.

The two Terminal-Bench figures come from different methods. Moonshot ran its model in the Claude Code harness and took each rival’s best published score; Artificial Analysis ran every model itself and puts K3.1 Pro first.

What does Kimi K3.1 Pro cost?

On Moonshot’s API, Kimi K3.1 Pro costs $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens. These are list prices in US dollars and exclude tax, which depends on where you are.

Sources: Kimi API pricing and Anthropic API pricing, both checked 31 October 2026. Both vendors’ prices exclude tax (Anthropic’s commercial terms, section H.2).

Every rate is a third below K3’s. The base input and output rates match Claude Sonnet 5.5 and are half of Claude Opus 5.5’s, although Sonnet 5.5’s cache hits cost half as much. Per-token prices only roughly compare across vendors, because tokenisers can split the same text into different numbers of tokens.

What you pay for a finished task depends on how much the model writes, and K3.1 Pro writes a lot. At launch, Artificial Analysis put it at $1.98 per index task, roughly the same as K3, even though its prices are a third lower, and found it used 55,000 output tokens per task, 14% more than K3. On version 5 of the index, which uses a different set of tests, its model page puts the cost at $3.68 a task.

The effort setting is the lever you control. Moonshot’s API defaults to max; Kimi Code defaults to high. In a small test published on launch day, Atomic Agent, which makes an open-source AI agent, ran one Python task, with five planted bugs and two missing features, three times per setting:

• at low effort, K3.1 Pro passed all 32 hidden tests in every run for $0.16 to $0.21 a task, taking 2.2 minutes on average
• at its default max effort, it cost $0.35 to $0.46 a task and took 4.9 minutes
• Claude Opus 5.5 passed the same tests for $0.37 a task in 69 seconds, and Claude Sonnet 5.5, in runs four days earlier, for $0.18 a task in 49 seconds

The top of each Kimi range allows for a billing gap Atomic Agent could not explain: its balance fell by $2.02 across the six Kimi runs, while the token counts came to $1.52. One task proves little, but the results fit Artificial Analysis’s token counts: turn the reasoning down before you judge the price.

Our own runs point the same way. We sent K3.1 Pro two synthetic jobs through OpenRouter, pinned to Moonshot’s own endpoint. One asked for 12 Google Ads headlines for a fictional Manchester boiler repair firm, checked by a script against seven rules, and ran three times per setting on 27 and 30 October 2026. The other was a budget-split puzzle with one right answer, run twice per setting on 27 October. At low or high effort each job cost about one US cent and took 16 to 41 seconds. At max, the ad brief cost 2.6 times as much as at low on average and the puzzle 1.7 times, and the slowest run took nearly two minutes. On the ad brief, leaving the effort unset cost about twice as much as low, in line with Moonshot’s documented default of max.

At every setting, each headline set passed all seven checks and each puzzle run found the right answer, though three ad runs and two puzzle runs per setting make a small sample. The checks cannot catch invented claims, and 10 of the 12 headline sets offered something the brief never mentioned, such as boiler servicing, Gas Safe registration, emergency call-outs or upfront quotes, so check every claim before an ad goes live.

A larger independent test shows the same trade-off. On Plotly’s data-analysis exam, 43 questions about a synthetic wind-farm database of 20 million rows, K3.1 Pro scored 42, up from K3’s 35, and found every planted pattern and data-quality problem. Its run cost $13.47 and took a median 2.5 minutes a question. OpenAI’s GPT-6.1 Sol scored 43 for $2.61, at about 18 seconds a question. Plotly ran each model once and says a gap of one or two questions is within the noise.

The API is pay as you go. K3.1 Pro becomes available once you have topped up at least $1, and the total you have topped up sets your rate limits. For work that can wait, the Batch API charges 60% of the standard price, although batch jobs still pay for cache writes, cache hits are not guaranteed and Zero Data Retention does not cover them.

In Kimi Code, K3.1 Pro needs Plus or above (Moderato on older plans), and its 1 million-token context and Highspeed need Pro or above (Allegretto on older plans). Highspeed also uses about twice the quota. Subscription quotas were a problem at launch: on 26 October Moonshot apologised for rate limits and unnecessary quota use by subscribers on K3.1 Pro, restored a quarter of every active subscriber’s monthly quota and promised changes so quotas would not run out so fast. Accounts opened from 31 October 2026 cannot use Moonshot’s older K2 models, and Moonshot plans to retire the series by the end of December 2026.

What are the drawbacks of Kimi K3.1 Pro?

Speed comes first. Artificial Analysis measures about 31 output tokens a second on Moonshot’s standard API, less than half the median of 81 for models in its price tier, and on its default workload of 10,000 input tokens the answer starts only after about 68 seconds, most of it spent thinking. Highspeed is Moonshot’s answer, at double the input and output rates, and on current measurements it works: Artificial Analysis puts it at about 70 tokens a second, with the answer starting after about 32 seconds. Our tests agree. We asked for the same 300-word product description at low effort, four times on each endpoint through OpenRouter on 27 October 2026 and again on 30 October, and Highspeed finished first in all eight pairs, in 10 to 14 seconds on average against 27 to 32. Kimi Code’s documentation adds that tool calls and scripts run no faster, so agent jobs that spend most of their time running commands gain less. Time it on your own work before paying for it.

The rest matter mainly if you build on the API:

• Moonshot says thinking cannot be switched off, and Kimi Code rejects requests that try. Through OpenRouter, though, our requests with reasoning turned off were answered on both 27 and 30 October 2026, with no reasoning tokens reported. Moonshot does not document that mode, so do not build on it.
• Temperature and top-p are fixed at 1.0 and 0.95, so you cannot adjust those sampling settings.
• Moonshot advises against using the API’s built-in web search in production for now, because it is being updated.
• Images must be sent as base64 data or uploaded files; the API does not accept public image URLs.
• In Kimi Code, a mistyped model ID falls back to the older K2.8 Preview without thinking, and switching models in the middle of a session degrades the output.
• The documentation is still settling. The English API quickstart has unfilled placeholders where its Terminal-Bench scores should be, and its hallucination figures differ from the launch post’s.

Is Kimi K3.1 Pro open source, and when can you download it?

Not yet, and perhaps not in the strict sense even when the weights arrive. Moonshot calls it open-weight and says the full weights will be released by 4 November 2026; until then Moonshot’s Hugging Face page has no K3.1 Pro repository, and Artificial Analysis lists the model as proprietary.

Moonshot’s launch materials do not say which licence the weights will carry. K3’s weights followed its launch by 11 days, under the Kimi K3 License, a custom licence Moonshot wrote itself. It grants MIT-style rights to use, modify and sell, with two conditions for very large commercial users:

• a business that runs what the licence calls a Model as a Service business, or has an affiliate that does, must sign a separate agreement with Moonshot before any commercial use if its revenue and its affiliates’ together top $20 million over any 12 consecutive months
• a commercial product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display “Kimi K3” prominently

Neither condition applies to internal use, which the licence defines as use that makes neither the model, its outputs nor its capabilities available to third parties, or to use through Moonshot’s own products and certified inference partners. Model as a Service means giving third parties access to language model inference or fine-tuning, for example through an API, with meaningful control over the inputs, parameters or training data. Products that only build the model into specific features, and services that merely relay requests to models hosted elsewhere, do not count. If K3.1 Pro arrives on the same terms, check both thresholds before you build a product or client service on it.

Open weights and open source are different things. The Open Source Initiative’s definition requires the freedom to use a model for any purpose without asking permission, along with the complete code used to train and run it and detailed information about the training data. A licence like K3’s, which makes the largest model-hosting businesses sign a separate agreement, would not meet it.

Running it yourself is likely to be a data-centre job. K3.1 Pro keeps K3’s 2.8 trillion parameters, and GPU cloud Vast.ai recommends a node of eight NVIDIA B300 GPUs to serve K3’s original 4-bit (MXFP4) weights. Our Muse Spark Open test covers what it takes to run a large open-weight model in practice.

Why did Moonshot cut cybersecurity data from Kimi K3.1 Pro?

Moonshot says more post-training kept making the model better at finding and exploiting software flaws, and that it does not want to release open weights that push the frontier of offensive hacking ability. It traced the gain to cyber-related material in its mid-training data and its reinforcement-learning tasks, and removed both for the open-weight release.

The launch post puts numbers on the effect. Moonshot’s internal xdev-bench-hard test asks a model to build working exploits for recently disclosed flaws in software such as the Linux kernel and Node.js. A version trained without those cuts solved 90.5% of its tasks within three attempts. The released K3.1 Pro solves 81.0%, against 33.3% for K3. Moonshot says that leaves it behind proprietary frontier models that have been public for at least three months, and argues that open weights help defenders audit code that rarely gets a security review.

For UK website owners, the jump from K3 matters more than the cap: 81% is far above 33%, and once the weights are out, copies cannot be recalled. Our Black October round-up covers October’s run of security disclosures, several of them found or sped up by AI, and what UK sites must do about them.

What should UK businesses check before using Kimi K3.1 Pro?

Check where your data goes and on what terms. Until the weights are out, the only hosted endpoints listed on OpenRouter and in Artificial Analysis’s provider comparison are Moonshot’s own. Moonshot AI is based in Beijing, but the privacy policy for its international platform names a Singapore company, Moonshot AI Pte. Ltd., as the provider and says data is stored on servers in Singapore. The model is not available through OpenRouter’s EU in-region routing either.

Moonshot’s API help centre says API inputs and outputs are not used to train or improve its models. Its platform privacy policy, last updated in April 2025, is broader. It says the prompts and files users submit help Moonshot optimise its models, and it lists “training and refining” them among its purposes, including in its section for UK and European users. It does not separate API customers from app users, so get the terms for your account in writing before sending client data.

Zero Data Retention, which deletes prompts and responses once a response is returned, is available to enterprise customers on request. It does not cover batch jobs, images or video sent through the file upload API, data handled by third-party models, connectors or plugins, or operational data such as security logs and billing records. Moonshot may also keep data where the law requires it. Kimi Business, which works on the web and in the desktop app but not yet on mobile, keeps organisation workspaces separate from personal accounts and says enterprise data is never used for training.

Kimi Work deserves the same scrutiny. The desktop app reads and handles your local folders, and at its most permissive setting, which Moonshot calls fully automatic, it can read and write files, run code and change system settings without asking. Its FAQ does not say where the model runs. A model of K3.1 Pro’s size will not run on a desktop computer, so expect what Kimi Work reads to be processed on remote servers, and ask Moonshot in writing which file contents leave your machine, and who processes them, before pointing it at client folders. Start with a test folder and the default or manual-approval setting.

Moonshot is also under US government scrutiny. In July, White House science adviser Michael Kratsios accused it, without publishing evidence, of distilling Anthropic’s Claude Fable 5 to build K3 and of using export-controlled NVIDIA chips. Treasury Secretary Scott Bessent warned that sanctions and Entity List designations would be “on the table” for Chinese firms running covert, industrial-scale distillation of American models. In September the NSA, FBI and CISA issued a joint advisory on what they called industrial-scale distillation campaigns by China-based AI companies; its public announcement names no company.

Two of Moonshot’s competitors have published accusations of their own. On 30 September OpenAI said it attributes the core of a July campaign to extract protected reasoning from its models to people associated with Moonshot, though it says it is unclear whether one actor was behind all of it. Anthropic’s September threat report says Moonshot silently passed some Kimi customers’ requests to Claude, showed them Claude’s answers and kept exchanges to train its own models. In one ten-day period, it says, Moonshot relayed almost 300,000 requests, and the rerouted requests included sensitive information such as internal code and live credentials. Anthropic says it does not know whether Moonshot told those customers.

Bloomberg reported on 28 October that Moonshot has not publicly addressed the allegations. Two people at the company had pushed back in July. On 21 July, the day before Kratsios spoke, Huang Zhenxin, who heads Moonshot’s enterprise business, told Daily Economic News (in Chinese) that K3’s leap in performance came from original architectural work and that the model was no distillation or copy of existing models. Two days later, the South China Morning Post reported, a Moonshot employee pointed to the short gap between the releases of Fable 5 and K3. Researchers are divided: some point to the lack of evidence, while others think distillation probably played a part.

For UK teams the practical questions are whether Moonshot’s service will stay available to you and where your requests go. Our post on the DeepSeek ban for US federal civilian agencies shows how one such restriction landed for another Chinese lab. If you plan to build on Moonshot’s API, ask in writing which companies process your requests, check the position again before you commit, and keep the option of moving to self-hosted weights once they appear and their licence suits you.

Should you switch to Kimi K3.1 Pro?

Not wholesale. It has earned a trial on the work where its independent scores are strongest, priced per finished task, and the promised weights will make it more interesting than it is today.

• If you run coding agents on long jobs, start here. Its Terminal-Bench 4.0 and AutomationBench-AA results from Artificial Analysis are the best case for it. Run it at low or high effort and log cost and time per completed task against your current model.
• If you do spreadsheet-heavy or financial analysis, the OfficeQA Pro v2 and finance results are promising, but they are Moonshot’s own numbers. Independent tests back the quality more than the speed: on Artificial Analysis’s AA-Briefcase test of agentic knowledge work it scores 1,676 Elo, behind Claude Sonnet 5.5 at maximum effort (1,814) and ahead of Gemini 4 Carbon at high (1,610), and on Plotly’s data-analysis exam it scored one question fewer than GPT-6.1 Sol but took far longer and cost five times as much. Check its outputs against the source documents before relying on them.
• If people wait on its answers, start with Highspeed and time whole replies. The standard endpoint streams at under a quarter of Claude Sonnet 5.5’s speed for the same base input and output rates (31 against 139 tokens a second on Artificial Analysis’s measurements), although Sonnet at maximum effort thinks for longer first: in Artificial Analysis’s end-to-end test, K3.1 Pro delivered 500 tokens in about 84 seconds against 467 for Sonnet. Highspeed doubles the input and output rates and, in our runs at low effort, answered a request for a 300-word description in 10 to 14 seconds on average.
• If client or regulated data is involved, wait for the weights and run it on infrastructure you choose, or get Zero Data Retention agreed in writing first, and run your usual UK GDPR checks either way.

Moonshot also launched 3D Make on kimi.ai and in the Kimi app, which turns a written prompt into a 3D model you can preview and edit online, download for 3D printing or export as a video.