What Is Kimi K3.1 Pro? Moonshot's 1M-Context Open-Weight Model: Price, How to Get It, the November 4 Weights, and What Its Benchmark Chart Does and Doesn't Show

What Moonshot shipped, in its own words

The launch post from @Kimi_Moonshot went up at 10:15 am PT on Thursday and had 805,000 views and 14,500 likes by evening. Its bullets: '1M-token context, 2.8T total / 104B active params (MoE), Multimodal'; 'Full-project coding: games in Unreal, Unity & Godot'; 'Excels at parametric 3D modeling in Blender & CAD'; 'More reliable at finance work: Accurate modeling with hallucination down by 65%'; 'Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.' The tech blog is more careful on two of those. The parity line there reads 'matches or approaches the strongest proprietary models,' and the hallucination figure is an internal stress test where claim-level hallucination 'fell from 23.4% for K3 to 8.3%' (a 64.5 percent relative drop; the blog never writes '65%'), with the caveat that 'the reported rates do not represent overall error rates in daily use.' The parameter count, notably, appears only in the X post; the blog and docs don't state it, so '2.8T / 104B' is inherited from K3's Hugging Face card (896 experts, 16 selected plus 2 shared). One more thing the blog says that the post doesn't: for the open-weight release Moonshot 'identified and removed cyber-related data from both training stages,' cutting its internal cyber benchmark from 90.5 percent on an uncleaned checkpoint to 81.0 on the shipped model, and says K3.1 Pro's 'cyber capability remains behind that of frontier proprietary models.'

Specs and limits

What it costs, and how to get it today

API. From Moonshot's pricing page: K3.1 Pro is $2.00 per million input tokens, $10.00 output, $0.20 for cache hits, with cache writes at $2 (five-minute) or $4 (one-hour). Highspeed doubles the base rates to $4/$20 with the same $0.20 cache hit. For context, K3 was $3/$15, so the new flagship is a third cheaper per token than the model it replaces. There are no long-context surcharges ('flat pay-as-you-go, with no tiering by context length'), access opens after a $1 top-up, and zero-data-retention is available for API customers. The same prices appear on Vercel's AI Gateway, which added both variants on launch day, and on OpenRouter, which lists a 1,048,576 context and image and video input but notes the model is 'not currently available through EU in-region routing.'

Consumer. Moonshot's international plans are Plus $19 a month, Pro $39, Max $99, and Ultra $199 (annual prices lower). In Kimi Code, Moonshot's CLI, VS Code extension, and desktop app, K3.1 Pro is 'Plus and above' with the full 1M context 'for Pro and above,' and Highspeed is 'Pro and above' at double quota usage. Kimi Work, 'a general-purpose local Agent for knowledge workers, built on Kimi Code,' needs version 3.1.0 or later. Enterprise admins must switch it on; 'access is off by default at launch.' Is it free? Moonshot's blog says K3.1 Pro 'rolls out today' on kimi.ai and the mobile apps, but no help-center page yet says whether free accounts can select it, and the help center's model list still shows K2.6 and K3. We'll update when that's documented; until then, assume a Plus plan. The weights. 'By Nov 4' is the only date. Nothing is on Hugging Face, and the license is unstated; K3 shipped under a modified MIT license with attribution and revenue clauses (UI credit above 100 million monthly users or $20 million monthly revenue), and most observers expect the same. Even then, this is not a model you run: K3's weights were about 1.4 TB with a validated minimum of sixteen GPUs, which is why Kimi sits in the also-considered list of our open-source ranking, a ranking of what you can actually run.

The benchmark chart, read with the footnotes

Moonshot's chart compares K3.1 Pro with Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1, DeepSeek V4.1 Flash, and GLM-5.3 across fourteen benchmarks, every model 'maxed out on thinking effort.' K3.1 Pro leads on six: SWE Marathon (51.7 vs 46.3 Opus, 42.5 Astra), Agents' Last Exam (43.2), OfficeQA Pro v2 (64.4), HLE with tools (70.0), GDP.pdf (34.6), and ZeroBench with tools (63.0). On the coding headline, Terminal-Bench 4.0, it scores 65.7, ahead of Astra's 58.2 and, in Moonshot's words, 'narrowly behind Claude Opus 5.5 (66.4%).' It trails clearly on Terminal-Bench Science (38.1 vs 68.1 Astra), on Moonshot's own long-horizon coding bench (61.2 vs 66.5 Astra), on BenchCAD with tools (92.7 vs 96.2 Opus) despite the 3D marketing, and on OSWorld (79.6 vs 81.8). Three footnotes matter. Kimi's Terminal-Bench and SWE Marathon runs used 'the Claude-Code harness,' while competitor numbers were 'taken as the maximum value reported in either their blog posts or the official leaderboard,' which is how Opus 5.5 shows 66.4 when the official tbench.ai board has it at 64.85. On OfficeQA, Astra was run through Codex and the others through Claude Code. And Gemini 4 Argon, currently #1 on Arena, is absent from the chart entirely. None of this makes the numbers wrong; it makes them Moonshot's numbers.

What independent runners say so far. Only one has scored it. Artificial Analysis posted within half an hour of launch: Intelligence Index 55, fourth place, behind Opus 5.5 at 58 and level with Sonnet 5.5 and Muse Wonder Full at 56; up eleven points from K3's 44. It also measured 67 percent on Terminal-Bench 4.0, which it called the highest of any model it has run, at $1.98 and about 55,000 output tokens per task. As of Thursday evening K3.1 Pro is absent from LiveBench, Arena, tbench.ai, SWE-bench Pro, and Vals. The packet's 'wait for the benches' crowd is right that the parity question is open; the AA number says the answer is probably 'close, not equal.'

Who's saying what

Bloomberg's framing was the company: 'Moonshot Debuts First AI Upgrade Since Its Own DeepSeek Moment,' calling it 'the first major upgrade to the seminal model that thrust the $50 billion company into the frontier,' a valuation Bloomberg reported on October 6 alongside an early-2027 IPO plan. The Next Web led with the cyber-data removal and recalled the summer's controversies: a White House official's July claim that Moonshot used banned Nvidia chips, and the distillation accusation around K3 that TechCrunch's experts dismissed. A Polymarket post ('claiming to outperform OpenAI's GPT-6 Astra on a key coding benchmark,' 595 likes) references no actual market; Polymarket's live 'best AI model end of October' contract had Moonshot at 0.15 percent on Thursday against Google's 87. Command Code, a terminal coding agent built for open models, had it live within ninety minutes on all plans, though its post rounds the parameters to '3T total.' Cursor, Windsurf, and the US inference hosts that carried K3 had not announced K3.1 Pro by Thursday evening.

Our read

Three things are true at once. K3.1 Pro is the strongest open-weight model anyone has measured, by a margin: 55 on Artificial Analysis is eleven points clear of its predecessor and within three of the best closed model. It is priced to be used, at $2/$10 against Opus 5.5's $4/$20 and Astra's $10/$50, with a million tokens of context and no long-context tax, and that price is the story for anyone running agents at volume. And the parity claim is still Moonshot's to prove: the two coding wins over Astra came through the Claude Code harness against leaderboard maximums, the independent boards haven't weighed in, and the weights everyone is excited about are a promise with a date and no license. If you build on models, try it today through the API or OpenRouter and judge it on your own tasks. If you run Kimi Code, it's in your Plus plan now. If you were waiting to download it, November 4 is the day to watch, and sixteen GPUs is the bill. We'll update this post when the weights land, when the license is published, and when the other leaderboards score it; the comparison table tracks the same.