A new name has joined the top tier of AI models, and it comes from China. On 23 October 2026, Beijing-based Moonshot AI launched Kimi K3.1 Pro, a huge AI model built for long, complicated jobs: writing and fixing code, working through financial documents, and building 3D models.
The headline numbers are eye-catching, but the honest story is more interesting: it beats GPT-6 on one important coding test, sits just behind Anthropic's best model overall, and is promised to become freely downloadable soon.
What is Kimi K3.1 Pro, in one minute?
Kimi K3.1 Pro is the latest version of Kimi, Moonshot AI's AI chatbot and model family. You can use it through the Kimi website and apps, a desktop app called Kimi Work, a coding tool called Kimi Code, and through Moonshot's API — the paid connection that lets other apps and developers use the model.
It is a workhorse for long tasks, not just quick questions. Moonshot says it is built for:
- Coding — from fixing a bug to building a whole project
- Agentic work — multi-step jobs where the AI uses tools, runs commands, and checks its own work
- 3D creation — describing a shape in code so it can be turned into a 3D model
- Financial research — working through real documents, spreadsheets, and reports
Importantly, K3.1 Pro is not built from scratch. Moonshot describes it as a post-trained version of Kimi K3 (July 2026) — the same base model, given a much bigger round of extra training on real tasks.
Who is Moonshot AI, and what is Kimi?
Moonshot AI is a Chinese AI company based in Beijing, founded in March 2023 and backed by major investors including Alibaba and Tencent. Kimi is its chatbot and model family, first released in October 2023. It grew quickly: Kimi K2 arrived in July 2025, followed by K2.5 and K2.6 in early 2026, then the much larger Kimi K3 in July 2026.
Kimi models are known for very long context windows and an "open-weight" approach, where the model files are eventually published so others can download and run them — unlike fully closed models such as GPT-6 and Gemini.
The 2.8-trillion number, explained simply
Kimi K3.1 Pro has 2.8 trillion parameters in total, with 104 billion active at any one time. A parameter is a tiny adjustable setting inside the model, tuned during training. More parameters generally mean a model can capture more patterns, but they also cost more computing power to run.
Kimi uses a mixture of experts design: imagine 896 specialist teams, of which only 16 are called in for any single question. Enormous total knowledge is available, but only a fraction works at each moment, keeping running costs down.
It also has a 1-million-token context window: the amount of text it can consider in one go. That is enough for a very large codebase or a stack of documents in a single session, which is why Kimi suits long jobs, not short chats.
If those terms are new, start here: Large Language Models Explained.
Does it really beat GPT-6?
Careful reading matters here. The short answer is: on one headline coding test, yes; overall, no — not yet, and not on every test.
Moonshot's own figures put Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, a tough test where an AI agent works in a computer terminal to complete realistic coding tasks. On the same test, Moonshot puts OpenAI's GPT-6 Astra at 58.2% and Anthropic's Claude Opus 5.5 at 66.4%. So Kimi beats GPT-6 Astra here, and lands just 0.7 points behind the current leader.
Other tests differ. On BenchCAD, a 3D-modelling test, Kimi scores 92.7% with tools, behind GPT-6 Astra (95.9%) and Opus 5.5 (96.2%). On Terminal-Bench Science, Kimi scores 38.1%, well behind GPT-6 Astra (68.1%). On Moonshot's hallucination test, Kimi improved sharply but still trails GPT-6 Astra.
Independently, the Artificial Analysis Intelligence Index gives Kimi 55, ahead of GPT-6 Astra and Gemini 4 Argon (both 53), but behind Sonnet 5.5 and Muse Wonder Full (56) and Opus 5.5 (58). Kimi ranks fourth of 215 models — the highest yet for a model set to become open-weight.
| Model | Artificial Analysis Intelligence Index | Terminal-Bench 4.0 (Moonshot's comparison) |
|---|---|---|
| Kimi K3.1 Pro | 55 | 65.7% |
| Claude Opus 5.5 | 58 | 66.4% |
| GPT-6 Astra | 53 | 58.2% |
| Gemini 4 Argon | 53 | 57.4%* |
* Gemini's Terminal-Bench figure comes from a separate comparison, not Moonshot's launch table, so treat it as indicative rather than like-for-like.
Caveat: most of Moonshot's coding scores were run inside Anthropic's Claude Code tool, while some rival scores came from leaderboards or the rivals' own reports, so conditions were not identical. For background on OpenAI's flagship, see GPT-6 and Intelligent UI Explained.
Price and availability: what is confirmed
Kimi K3.1 Pro is available now through Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code, and the Kimi API. On business plans, an administrator must enable it — it is off by default at launch. A faster sibling, Kimi K3.1 Pro Highspeed, offers the same abilities at two to three times the speed.
For developers, Moonshot lists $2 per million input tokens and $10 per million output tokens — down from K3's $3 and $15. Highspeed costs double; cached input on the standard model costs $0.20 per million. These are developer prices, not a consumer subscription, and third-party resellers differ.
The catches to know before getting excited
- The open weights are a promise, not yet a fact. Moonshot says the full model weights will be published by 4 November 2026, and the licence governing them had not been published at launch. Until then, you can only use Kimi through Moonshot's own app and API — you cannot yet download and run it yourself.
- Moonshot deliberately reduced one capability. It removed cyber-related data from training before the open-weight release; on its internal test, that cut the vulnerability-research score from 90.5% to 81.0%.
- Benchmarks are not the whole story. Kimi trails badly on science tasks, and its hallucination rate is still above GPT-6 Astra's on Moonshot's own test. Verify important outputs before relying on them.
What it means for ordinary users
You do not need to be a programmer to feel the effect. A near-frontier model at $2/$10 pricing, soon downloadable, pressures rivals on price and quality — which eventually reaches everyday apps. And if models can handle multi-step jobs, AI moves from answering questions to finishing tasks.
Practical move: try Kimi in the app for a long document or coding question, and wait until after 4 November before judging the open-weight claim.
Takeaway checklist
- Kimi K3.1 Pro launched 23 October 2026 — not 25 October, as some early summaries suggested.
- 2.8 trillion total parameters (104 billion active), 1-million-token context window, mixture-of-experts design.
- Beats GPT-6 Astra on Terminal-Bench 4.0 (65.7% vs 58.2%), but trails Claude Opus 5.5 (66.4%) and GPT-6 Astra on 3D and science tests.
- Independent ranking: 55 on the Artificial Analysis Intelligence Index, fourth overall, ahead of GPT-6 Astra and Gemini 4 Argon (53 each).
- API pricing: $2 input / $10 output per million tokens, Highspeed variant at double the rate.
- Full open weights promised by 4 November 2026 — until then app- and API-only, licence still unpublished.
- Try it for long coding, document, or 3D jobs; verify important outputs, and wait for the weights release before calling it the best open model available.
