Moonshot AI launched Kimi K3.1 Pro on 23 October 2026, in posts on its X account and a tech blog. It calls the model “our most capable 3T-class open-weight model”, with 2.8T total parameters, 104B active, and a 1-million-token context window.
The API price, in an image Moonshot posted on X, is $2.00 per million input tokens and $10.00 per million output tokens. That is half of the $4 and $20 Anthropic lists for Claude Opus 5.5, by our arithmetic. Moonshot says the model weights will be released by 4 November.
Business Pill · OPEN WEIGHTS EXPLAINED
A short Business Engineer explainer of open weights: rent the model or run it yourself. It teaches the general idea only and says nothing about Moonshot or Kimi K3.1 Pro.
The key insight: As we read it, the new element is where Moonshot set the price: $2 and $10 per million tokens, the same list price Anthropic gives Claude Sonnet 5.5 and half of Claude Opus 5.5, for a model its blog places just behind Opus 5.5 on Terminal-Bench 4.0.
Update (23 Oct): how Moonshot says it limited cyber capability before the open-weight release. In the cybersecurity section of its tech blog, Moonshot says it found cyber-related content in its general mid-training corpus and cyber-related tasks in its reinforcement-learning mix, and that “For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness.” On its in-house vulnerability benchmark, Moonshot says the checkpoint without these measures “solves 90.5% of problems at pass@3; the final K3.1 Pro reaches 81.0%.” It adds that “K3.1 Pro’s cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months.” These are Moonshot’s own internal results; we have not seen an independent evaluation.
Update (23 Oct): Tencent posts an open interactive world model the same week. Tencent’s Hunyuan team has published Hy-Nomas on Hugging Face under an Apache 2.0 licence (the model page was created on 21 Oct and updated on 23 Oct). Its model card calls it “a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control” that “generates a world one video segment at a time.” Per the card, the distilled setup uses 8 denoising steps for 720 × 1280 output at 24 FPS; the standard setup uses 8-way sequence parallelism across eight GPUs, and a separate setup runs on “a single NVIDIA GeForce RTX 4090” using INT8 weights and CPU offload. Tencent also released Hy-Nomas-Bench, 280 test cases; the card lists the technical report as “coming soon” and gives no benchmark results, so there is no performance claim to compare yet.
Update (23 Oct, evening): a first third-party ranking. Artificial Analysis, which runs its own independent index of AI models, posted on X that “Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published.” The post says the new model “gains 11 points over Kimi K3”. Its model page says the Intelligence Index v4.3.2 combines 10 evaluations, including Terminal-Bench 4.0 and Humanity’s Last Exam. This is Artificial Analysis’s own index, not Moonshot’s; we read the post and page on 23 October at about 20:00 UTC.
Update (23 Oct): another lab is already shipping on top of a Chinese open-weight model. humans& released Ampersand v0.1 on 20 October, and its launch post says: “We trained Ampersand from GLM 5.3, released by Z.ai.” The company describes Ampersand as “a collaborator trained to help people achieve their individual and collective goals” and says it is available “through our Agents API, with no waitlist.”
Its pricing page lists $1.00 per million input tokens and $5.20 per million output tokens. In its post on X, humans& says that in multi-month simulations users collaborating with Ampersand accomplished more than with Fable and other frontier models; that is the company’s own evaluation.
The model. Moonshot’s X post gives 2.8T total and 104B active parameters in a mixture-of-experts design, a 1M-token context and multimodal input. The tech blog says K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute.
Two versions. Kimi K3.1 Pro Highspeed, according to the blog, delivers 2-3ร the speed of K3.1 Pro. Both are live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, the X post says.
Open weights later. The blog says: “The full model weights will be released by Nov 4.” On 23 October the newest model on Moonshot’s Hugging Face page was still Kimi-K3.
List output price per million tokens. Kimi K3.1 Pro: Moonshot’s pricing image on X, 23 October 2026. Claude models: anthropic.com/pricing, read 23 October 2026. MiniMax M3.1 Flash: MiniMax’s listing as reported in our 19 October piece.List output price per million tokens: Claude Fable 5.1 $50, Claude Opus 5.5 $20, Kimi K3 $15, Kimi K3.1 Pro $10, Claude Sonnet 5.5 $10, MiniMax M3.1 Flash $0.60. Sources: Moonshot’s X post and platform.kimi.ai, anthropic.com/pricing, MiniMax’s listing as reported in our 19 October piece.
What Was Posted on X
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
๐น 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. ๐น Full-project coding: games in Unreal, Unity & Godot. ๐น Excels at parametric 3D modeling in Blender & CAD. ๐น More reliable at finance work: Accurate modeling with hallucination down by 65%. ๐น Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2โ3ร faster) are now live on http://Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Moonshot AI introduced Kimi K3.1 Pro on X as its most capable 3T-class open-weight model, with a 1M-token context and 2.8T total and 104B active parameters, and said open weights will come by Nov 4, 2026.
Bloomberg wrote on X that Moonshot AI, which it called the $50 billion company and a Chinese upstart that had its own DeepSeek moment over the summer, released the first major upgrade to the model that thrust it into the frontier.
Kimi K3.1 is comparable to Claude opus 5.5 and OpenAIโs GPT 6 Astra!!! We are getting really good frontier open source models!
An X user wrote that Kimi K3.1 is comparable to Claude Opus 5.5 and OpenAI’s GPT 6 Astra, calling it a sign of really good frontier open-source models; the claim is the poster’s own.
INTELLIGENCE is becoming too cheap to meter
for example, last year opus 4 cost $30 per million tokens with an AA intelligence score of 21. now opus 5.5 costs just $8 with a score of 58
and with openai, gpt-5 cost $3.44 with a score of 23 while 6.1 sol now costs $4 with a score of 52
Haider argued on X that intelligence is becoming too cheap to meter, citing figures of their own: Opus 4 at $30 per million tokens with an Artificial Analysis score of 21 a year ago versus Opus 5.5 at $8 with a score of 58, and GPT-5 at $3.44 with a score of 23 versus GPT-6.1 Sol at $4 with a score of 52; the numbers are the poster’s own.
The Price Table
Moonshot’s pricing image says “K3.1 Pro launches at 2/3 of K3’s price”. Kimi K3 is listed at $3.00 input and $15.00 output on platform.kimi.ai. The table compares list prices from each company’s own price list, per million tokens.
Model
Input
Output
Source
Kimi K3.1 Pro
$2.00
$10.00
Moonshot X post, 23 Oct
Kimi K3.1 Pro Highspeed
$4.00
$20.00
Moonshot X post, 23 Oct
Kimi K3
$3.00
$15.00
platform.kimi.ai
Claude Fable 5.1
$10
$50
anthropic.com/pricing
Claude Opus 5.5
$4
$20
anthropic.com/pricing
Claude Sonnet 5.5
$2
$10
anthropic.com/pricing
MiniMax M3.1 Flash
$0.15
$0.60
MiniMax listing, as in our 19 Oct piece
Our arithmetic: K3.1 Pro is half of Opus 5.5 on both input and output, the same as Sonnet 5.5, and one fifth of Fable 5.1. MiniMax M3.1 Flash is about one thirteenth of K3.1 Pro on input and one seventeenth on output.
What Moonshot Claims on Benchmarks
These are Moonshot’s own results, published in its tech blog. On Terminal-Bench 4.0, the blog says “K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%)”. Its notes say K3.1 Pro was evaluated with the Claude-Code harness.
3D and CAD. On BenchCAD the blog gives 92.7% with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.
Finance. The blog says K3.1 Pro scored 64.4 on OfficeQA Pro v2, which it calls state-of-the-art.
Hallucinations. On its internal Kimi Hallu Bench Hard, the blog says “its claim-level hallucination rate fell from 23.4% for K3 to 8.3%”. It adds: “The reported rates do not represent overall error rates in daily use.”
On cybersecurity, the blog says: “K3.1 Pro’s offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models.”
Connecting the Coverage
Our MiniMax pricing piece reported on 19 October that MiniMax lists M3.1 Flash at a flat $0.15 per million input tokens and $0.60 per million output tokens. Our MiniMax weights piece reported that MiniMax published the M3.1 Flash weights on 23 October. Our export filing piece reported that DeepSeek and MiniMax completed registration under China’s new pre-export filing rule for AI technology. Our MiniMax and Zhipu piece covered the two companies’ share prices falling from their peaks.
The Structural Read
As we read it, the step from Kimi K3 is a price cut and a capability claim at once: the pricing image says “K3.1 Pro launches at 2/3 of K3’s price”, and the blog says K3.1 Pro is a post-trained version of Kimi K3.
As we read it, the open-weight promise is the other half of the offer: the blog says the full model weights will be released by Nov 4, and the posts and blog we read do not state the licence they will carry.
As we read it, the gap to MiniMax is wide: MiniMax lists M3.1 Flash at $0.15 and $0.60, so Moonshot is pricing nearer to Anthropic’s mid-tier than to the cheapest Chinese open model in our table.
Moonshot AI, Kimi K3.1 Pro tech blog, 23 October 2026
“K3.1 Pro achieves consistent gains in every domain”
Three Implications
FOR DEVELOPERS The X post says K3.1 Pro and K3.1 Pro Highspeed are live on the Kimi API, and the blog says Highspeed delivers 2-3ร the speed of K3.1 Pro.
FOR ANTHROPIC AND OPENAI Moonshot’s blog compares K3.1 Pro with Claude Opus 5.5 and GPT-6 Astra on its own benchmark runs, including Terminal-Bench 4.0 and BenchCAD.
FOR OPEN-WEIGHT HOSTS The blog says the full model weights will be released by Nov 4; the licence was not stated in the posts and blog we read.
The framework says the winning pattern is: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Moonshot is opening the model layer, with weights promised by 4 November, while selling the API and the Highspeed version as the paid layer. The framework was published in March 2026 and does not discuss Moonshot or Kimi K3.1 Pro.
What Is Not Established
The benchmark results in Moonshot’s launch posts are Moonshot’s own. The first independent result we found, later on 23 October, is Artificial Analysis’s Intelligence Index score in the Update above; it is a single index, not a review of Moonshot’s own benchmark table.
The licence for the open weights is not stated in the X posts or the tech blog we read. The weights were not yet on Hugging Face on 23 October.
The K3.1 Pro prices come from the image in Moonshot’s X post. The pricing page we read on platform.kimi.ai listed Kimi K3 but not yet K3.1 Pro.
We did not read OpenAI or DeepSeek price lists for this piece, so their models are not in the table.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This piece is part of this week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Moonshot has put a 2.8T-parameter model on its API at $2 and $10 per million tokens, half of Claude Opus 5.5’s list price, and says the weights will be released by 4 November. Its blog places K3.1 Pro just behind Opus 5.5 on Terminal-Bench 4.0, a result Moonshot measured itself.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece is based on Moonshot AI’s three posts on X and its tech blog of 23 October 2026, read in full, the pricing image in its second post, the Kimi API and Anthropic pricing pages, and our own earlier pieces. The benchmark figures are Moonshot’s own. We did not contact Moonshot. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
