TECH

Moonshot AI Launches More Powerful Kimi Model

Published: Oct. 26, 2026 11:44 a.m. GMT+8

00:00

00:00/00:00

Listen to this article 1x

Moonshot AI launched Kimi K3.1 Pro, its most powerful model yet. Photo: VCG

Chinese AI startup Moonshot AI launched its most powerful model yet, cutting prices while boosting performance in coding and agent tasks as domestic developers try to keep pace with a rapid series of upgrades from U.S. rivals.

The Beijing-based company, which operates the Kimi chatbot, released Kimi K3.1 Pro early Saturday. Moonshot said the model has about 3 trillion parameters, supports a context window of 1 million tokens and shows particularly strong gains in coding, 3D creation, financial research and long-running agent tasks.

You've accessed an article available only to subscribers

Subscribe today for just $.99.

VIEW OPTIONS

Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.

Subscribe to both Caixin Global and Bloomberg - for the price of one.

Share this article

Open WeChat and scan the QR code

DIGEST HUB

Digest Hub

Back

Explore the story in 30 seconds

Moonshot AI launched Kimi K3.1 Pro (≈3T parameters, 1M context), scoring 55 on the AI Index (4th), behind Anthropic’s Opus 5.5, Sonnet 5.5, and Meta’s Muse.
API prices dropped from K3: input 1.3 yuan cached, 13 uncached, 68 output (previously 2, 20, 100).
Release follows U.S. rival upgrades; Moonshot pushes enterprise sales, having raised ~$8.5B with a $50B pre-IPO valuation.

AI generated, for reference only

Who’s Who

Moonshot AI: Moonshot AI is a Beijing-based Chinese AI startup operating the Kimi chatbot. It launched Kimi K3.1 Pro, a post-trained model with ~3 trillion parameters and 1M token context, excelling in coding, financial research, 3D creation, and agent tasks. API prices were cut. Valued at $35B, Moonshot is expanding enterprise services and preparing for a Hong Kong listing.

Anthropic: Anthropic is a U.S. AI developer mentioned as a competitor to Chinese startups. Its models Opus 5.5 and Sonnet 5.5 scored 58 and 56 on the Artificial Analysis Intelligence Index, ranking above Moonshot's Kimi K3.1 Pro (55). Kimi scored above Anthropic's Fable 5.1. Since September, Anthropic released new models, with its frontier models also ranking higher on Code Arena.

Meta: The article mentions Meta in two contexts: its "Muse Wonder Full" model, which ranks above Moonshot's K3.1 Pro on the Artificial Analysis Intelligence Index, though no public API is available for it. Meta also introduced new models since September, alongside Anthropic, OpenAI, Google and SpaceXAI, and its frontier models rank highly on coding benchmarks like Code Arena.

OpenAI: The article mentions OpenAI's GPT-6 Astra, which scored below Moonshot AI's Kimi K3.1 Pro on the Artificial Analysis Intelligence Index. OpenAI is also cited among U.S. developers (with Anthropic, Google, Meta, SpaceXAI) that introduced new models since September, and its frontier models ranked above K3.1 Pro on Code Arena.

Google: According to the article, since September, Google introduced new models, among other U.S. developers. Additionally, Moonshot held talks with Microsoft, Amazon, and Google about hosting its K3 model on their cloud platforms. No further details about Google are provided.

SpaceXAI: SpaceXAI is mentioned as one of the U.S. developers that introduced new models since September, alongside Anthropic, OpenAI, Google, and Meta. No further details are provided in the article.

Zhipu AI: According to the article, Zhipu AI (智谱AI) makes the GLM-5.3 model. Its average cost per task is roughly in line with Moonshot's K3.1 Pro, which was $2.11 in Artificial Analysis testing. No other details about Zhipu AI are mentioned in the article.

Alibaba: According to the article, Alibaba is referenced as the developer of the Qwen 3.8-Max model. In Artificial Analysis testing, Moonshot AI's K3.1 Pro had an average cost per task of $2.11, which was below Alibaba's Qwen 3.8-Max, indicating Alibaba's model was more expensive on that measure.

Microsoft: The article mentions that Moonshot AI held talks with Microsoft, Amazon, and Google about hosting its K3 model on their cloud platforms. No further details about Microsoft are provided in the article.

Amazon: According to the article, Amazon held talks with Chinese AI startup Moonshot AI about hosting its K3 model on cloud platforms. Amazon has since made K3 available through Amazon Bedrock and its Kiro coding assistant, expanding access to Moonshot's models.

Artificial Analysis: According to the article, Artificial Analysis is a platform that evaluates AI models. It runs an Intelligence Index where Moonshot's K3.1 Pro scored 55, ranking fourth, trailing Anthropic's models and Meta's Muse Wonder Full. It also measures average cost per task ($2.11 for K3.1 Pro) and hosts benchmarks like Code Arena.

AI generated, for reference only

What Happened When

July 2026：: K3.1 Pro's predecessor K3 topped the Code Arena coding benchmark, becoming the first Chinese model to do so.

July 2026：: Moonshot launched K3 and introduced AI products for legal and financial applications, and started a partner program with IT service providers and system integrators.

July 2026：: Moonshot completed a Series F funding round at a post-money valuation of $35 billion, led by China's national AI investment fund and other state-backed investors.

July 2026：: Following K3's launch, Moonshot temporarily suspended new subscriptions.

August 2026：: Moonshot and its partners shared revenue based on token usage, a model in place since this month.

August 2026：: Moonshot began raising a pre-IPO round at a valuation of $50 billion before new investment.

End of August 2026：: Moonshot had raised nearly $8.5 billion in private markets and had shortened its potential timetable for a Hong Kong listing to as little as six months, potentially targeting early 2027.

Since mid-2026：: Moonshot pursued a broader commercial business strategy.

Since 2025：: Moonshot increasingly sought enterprise customers.

September 2026：: Amazon made K3 available through Amazon Bedrock and its Kiro coding assistant, after talks with Microsoft, Amazon, and Google about hosting the model.

Since September 2026：: Anthropic, OpenAI, Google, Meta, and SpaceXAI introduced new models, and U.S. developers rolled out cheaper offerings.

Before October 4, 2026：: K3 had cost 2 yuan per million input tokens when cached, 20 yuan when uncached, and 100 yuan per million output tokens.

Early October 2026：: K3.1 Pro showed improvement on some third-party benchmarks, as measured in testing conducted during this period.

Early October 2026：: K3.1 Pro's average cost per task was $2.11 in Artificial Analysis testing, roughly in line with Zhipu AI's GLM-5.3 and below Alibaba's Qwen 3.8-Max.

October 4, 2026：: Moonshot AI launched Kimi K3.1 Pro early Saturday, releasing its most powerful model with about 3 trillion parameters, a 1 million token context window, and strong performance gains in coding, 3D creation, financial research, and long-running agent tasks.

October 4, 2026：: Moonshot cut API prices for K3.1 Pro from those charged for K3, with costs of 1.3 yuan per million input tokens when cached, 13 yuan when uncached, and 68 yuan per million output tokens, effective from its release.

October 4, 2026：: K3.1 Pro ranked eighth on Code Arena after its launch, behind frontier models from Anthropic, OpenAI, and Meta.

October 4, 2026：: The release followed nearly a month without a new flagship model from a major Chinese developer.

October 2026：: K3.1 Pro's capabilities in financial research, 3D modeling, and cybersecurity were highlighted in Moonshot's release materials.

Week of October 5-11, 2026：: K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, ranking fourth on the platform, trailing Anthropic's Opus 5.5 (score 58) and Sonnet 5.5 (score 56), and Meta's Muse Wonder Full.

AI generated, for reference only