It launched on 23 October with a faster Highspeed version, and Moonshot’s launch post says the full model weights, the files needed to run it yourself, arrive by 4 November. It is live in the Kimi app, at kimi.ai, in Kimi Work, in Kimi Code and through the Kimi API, the paid route for developers.
Benchmarks show how it does on someone else’s exam, so the prompt below runs one of your real tasks through Kimi and through the AI you use every day, then gives you a blank scorecard for each answer.
The prompt: run one real task through both
You are helping me find out whether a new AI model can do one of my real tasks as well as the AI I already use. I am pasting this exact prompt into two different AI tools, so treat it as a fair exam: same task, same rules, no extra effort for either side. If you pad, guess or hide a weak spot, I may switch tools or keep paying for the wrong one, so a plain answer I can check matters more than an impressive one. ABOUT MY TEST (I fill this in once; any line can say "not stated") - The task, in my own words: [e.g. "turn these meeting notes into a one-page summary with actions and owners", "reply to this customer complaint", "find the mistake in this spreadsheet formula", "plan three days in [city] for two people on a budget of [amount]"] - The material for the task: [paste it below this list, with names, numbers and client details swapped for placeholders like [CLIENT], [AMOUNT], [COLLEAGUE], or write "none"] - What a great answer must include: [e.g. "every action has an owner and a date", "under 200 words", "a formula that works in Google Sheets", "costs for every day"] - What would make me bin the answer: [e.g. "any invented fact", "a stiff, robotic tone", "ignores the budget", "longer than one page"] - What I already know is the right answer, if anything: [e.g. "the total should be 1,240", "the meeting agreed to delay the launch", "not stated"] - How I will judge it (up to five points; if blank, use accuracy, follows my instructions, complete, tone, usable without edits): [e.g. "accuracy, tone, length"] WORK IN STAGES Stage 1. Privacy check, before anything else. Read my task and material. If you see anything that looks like real personal or confidential data (a full name with contact details, an email address, a phone number, a home address, an account, card or ID number, health details, a password or key, or anything marked confidential), stop. List those lines, suggest a placeholder for each, and do not do the task until I reply "go". If you see nothing like that, write "Privacy check: nothing flagged" and carry on. Stage 2. Check you understand. In three lines or fewer, restate the task as you understand it. If something you need is missing and the answer depends on it, ask me up to three questions and stop. If what is missing is minor, write [NEED: what] at that point, make the smallest sensible assumption, label it "Assumption:", and carry on. Stage 3. Do the task. Give your best answer once, in the format I asked for. If I gave no format, use the shortest format that does the job. Do not invent facts, figures, quotes, sources, links or prices. Where you would need one you do not have, write "not stated" or [NEED: what] in its place. If I told you the right answer and your working disagrees with it, say so plainly; do not quietly bend your answer to match mine. Stage 4. Show me what to check. Under the heading "Check these", list: - every fact, number or named thing in your answer, and where it came from: my material, general knowledge, or an assumption; - anything you were unsure about, and how unsure; - anything in my instructions you could not follow, and why. If you used no outside facts at all, say so. Stage 5. Give me a blank scorecard. Print a table for me to fill in by hand, with my judging points from the set-up as the rows. Columns: "Point", "Score 1 to 5", "Note". Add two rows at the end: "Edits I had to make before I could use it" and "Would I use this as it is? yes / with edits / no". Leave every score empty. Do not score yourself. SELF-CHECK BEFORE YOU SEND - You ran the privacy check first, and stopped if anything was flagged. - Every gap is marked "not stated", [NEED: what] or "Assumption:", with nothing filled in by guessing. - The "Check these" list covers every number and named fact in your answer. - The scorecard uses my judging points and every score is blank. - You did not mention, rate or compare any other AI tool.
Run the test, step by step
- Pick a task you already know the answer to. Notes you summarised last week, a formula you have already fixed or an email you have already sent all work, because you can spot a wrong answer without checking anything else.
- Swap private details for placeholders before you paste. Kimi’s privacy policy says your personal information is stored in China and chats can be used to improve its models, so replace names, numbers and client details with tags like [CLIENT] and [AMOUNT].
- Run the prompt in a fresh chat in your usual AI first. Its privacy check then catches anything you missed before the same text goes to Kimi, and a new chat stops what that AI remembers about you from giving it an edge.
- Open Kimi, choose K3.1 Pro and paste the same text. Sign in at kimi.ai or in the Kimi app, pick K3.1 Pro from the model picker, and paste the identical wording and material with no extra hints. Moonshot’s privacy policy says sign-up may ask for a mobile number, or you can log in with another account such as WeChat.
- Time both answers, from your last reply to the finished answer. Plotly’s launch-day test found K3.1 Pro took a median 2.5 minutes per question, the second slowest of 34 models. Plotly sent data-analysis questions through the API, so your wait may differ. If speed matters most, Highspeed (2 to 3 times faster, says Moonshot) needs the Pro plan or above.
- Fill in each scorecard straight after reading that answer. Check every line under “Check these” against what you already know, then score, so the second answer cannot colour how you rate the first.
- Repeat with two different kinds of task before you decide. One task can flatter either tool, so add a writing job, a numbers job or a planning job. If the totals tie, keep the tool you already use, because switching costs you time.
What the scores say so far
Most of the numbers in the launch post are Moonshot’s own, so the independent ones come first.
- Independent, Artificial Analysis. Its model page gives K3.1 Pro 55 on its Intelligence Index, a combined score across the tests it runs on every model, which puts it fourth among comparable models, where the median is 27. The same page calls it notably slow and says it writes very long answers.
- Independent, Plotly. On Plotly’s private data-analysis exam it passed 42 of 43 questions, tied third of 34 models and up from 35 for Kimi K3. The whole exam cost $13.47, paid per use through OpenRouter, where GPT-6.1 Sol scored 43 for $2.61; a chat subscription is priced differently.
- Moonshot’s own, coding. On Terminal-Bench 4.0, a public coding test, Moonshot ran K3.1 Pro itself and reports 65.7%. The 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra come from other published results, so the setups may differ.
- Moonshot’s own, made-up facts. On Kimi Hallu Bench, a hard internal test built from prompts chosen to trip models up, the share of false claims fell from 23.4% for K3 to 8.3%. Moonshot says these rates do not represent everyday error rates.
What to check before you paste anything
- Where app chats are stored. Moonshot’s privacy policy for the Kimi website and app, last updated 24 August 2026 and read 26 October 2026, names Beijing Moonshot Technology as the company and says: “We store your personal information within the territory of the People’s Republic of China.”
- Training on your chats. The same policy says de-identified inputs and outputs are used to optimise its models. To opt out, email support@moonshot.cn; Kimi’s help page says it takes 5 to 7 working days and covers your whole account, past chats included.
- Memory. The policy says Kimi may remember things you mention to shape later replies, and that you can switch this off and review what it saved in the app’s settings.
- Where API data is stored. Moonshot’s API privacy policy, last updated 30 April 2025 and still the version posted when read 26 October 2026, says the API is run by Moonshot AI PTE and data sits on servers in Singapore. Its API terms, updated 30 July 2026, say Moonshot may use your content to improve its services.
- Get API promises in writing. The launch post says Zero Data Retention, an option where Moonshot keeps no copy of what you send, is available to API customers, and Kimi’s ZDR page says to enterprise customers, so ask for it in writing before you send work data.
- Check who trains on API data. Kimi’s API data security page and its API terms give different answers on whether API data trains its models.
- Open weights are coming, licence unknown. Once the weights are out, a company can run K3.1 Pro on its own servers and send Moonshot nothing. Plotly notes the licence is not published yet and that K3’s weights were about 1.6 TB, which needs a powerful specialist server, well beyond a laptop.
The honest bit
Moonshot’s K3.1 Pro pricing page, read 26 October 2026, says the free plan includes limited K3.1 Pro on the web and in Kimi Work, but not how much. If the free allowance runs out partway through your test, write that on the scorecard, because it is part of what the free plan gets you.
The facts here come from Moonshot’s launch post and pricing page, Kimi’s privacy policy and help pages, the Kimi API privacy policy and terms, Artificial Analysis and Plotly, read 26 October 2026.
Do it now
Pick one task you finished last week, swap out the names, and run it through both tools tonight. Two filled-in scorecards will tell you whether Kimi earns a tab next to the one you already have open.
A few quick questions
Is Kimi K3.1 Pro open source?
Not yet. Moonshot says the full weights arrive by 4 November 2026, and until then Artificial Analysis lists it as proprietary. Plotly reports that the licence has not been published, so read it before you build anything on the model.
What does the Kimi API cost for K3.1 Pro?
Artificial Analysis lists $2.00 per million input tokens and $10.00 per million output tokens on Kimi’s own API. A token is a small chunk of text, often part of a word, and you pay for what you send and what comes back. It writes long answers, and Plotly found output was the biggest part of its bill.