Moonshot's Kimi K2: Alibaba-Backed Open Model Outscores ChatGPT and Claude in Coding

Moonshot AI, a Beijing startup backed by Alibaba, released Kimi K2 in mid-July 2025, and the model quickly became one of the most discussed Chinese entries in the generative AI race. Its pitch rests on three points: strong coding performance, open availability, and a price far below the leading US models.

That combination is familiar. Low cost and open source were the same two ingredients that made DeepSeek's January 2025 release so disruptive. Open-source models make their weights freely available, an approach that few large US tech companies have embraced, with Meta and, to a degree, Google as the main exceptions.

The timing was notable. Within hours of the Kimi K2 launch, OpenAI chief executive Sam Altman said the company's first open-weight model would again be delayed indefinitely while it carried out more safety testing.

Benchmarks: Where Kimi K2 Claims the Lead

Coding is the headline capability. Writing software is one of the areas where businesses see the clearest potential for generative AI to augment or replace work, which is why Anthropic made coding the focus of Claude Opus 4 when it launched in late May 2025.

In its launch posts on X and GitHub, Moonshot said Kimi K2:

• outperformed Claude Opus 4 on two benchmarks,
• delivered better overall results than OpenAI's coding-oriented GPT-4.1 across several industry measures.

Independent analysts were broadly positive but measured. Wei Sun, principal AI analyst at Counterpoint, described the model as globally competitive and open source. She also cautioned that Moonshot is still building out the tooling needed to plug Kimi into other systems. Until that ecosystem matures, developers may find it hard to move away from rival platforms, which makes the raw coding gains less of a game changer than the benchmark numbers alone suggest.

Pricing That Undercuts US Rivals

Price is where Kimi K2 stands out most. The model can be used for free through Kimi's app and web interface, while ChatGPT and Claude charge monthly subscriptions for their premium tiers. API pricing at launch compared like this:

That makes Claude Opus 4 around 100 times more expensive for input and 30 times more expensive for output. Tokens are the units in which models measure the text they process, so these differences compound quickly at scale. Analysts singled out the low token costs as a strong draw for large deployments and budget-conscious teams.

License Terms and Early Reception

The license is permissive. Developers can use K2 largely as they wish. The main condition applies only to very large commercial products: if a service built on the model exceeds 100 million monthly active users or $20 million in monthly revenue, it must display "Kimi K2" in its user interface.

Early feedback on social media, in both English and Chinese, was mostly positive. Some users reported hallucinations, the familiar problem of models inventing information. Others were enthusiastic: Pietro Schirano, founder of the AI design startup MagicPath, wrote that K2 was "the first model I feel comfortable using in production since Claude 3.5 Sonnet."

China's Place in the Global AI Race

Moonshot has released open models before. Its chatbot gained popularity in early 2024 as an alternative to ChatGPT, which is not officially available in China. The domestic market has since become crowded, with similar chatbots from ByteDance and Tencent and a search engine overhaul with AI features from Baidu.

Kimi K2 arrives as investors increasingly look at Chinese alternatives to US AI technology. The landscape around it was shifting at the same time:

• DeepSeek had not yet shipped a major update to its R1 and V3 models despite the excitement around its earlier releases.
• Manus AI, another Chinese startup that drew attention earlier in 2025, moved its headquarters to Singapore.
• OpenAI had not yet released GPT-5.

Counterpoint's Sun suggested that work on GPT-5 might be absorbing OpenAI engineering capacity that could otherwise go to its open model. She added that releasing a strong open model is difficult without eroding the competitive advantage of a company's proprietary models.

Kimi-Researcher and the Grok 4 Comparison

K2 was not Moonshot's only recent release. A month earlier the company launched a Kimi research model and claimed it matched the 26.9 score of Google's Gemini Deep Research on the Humanity's Last Exam benchmark while beating OpenAI's equivalent.

The research model even came up during the launch of xAI's Grok 4. On the same benchmark, Grok 4 scored 25.4 on its own and 44.4 when it was allowed to use tools and web search.

Winston Ma, an adjunct professor at NYU School of Law and author of a book on China's tech power, described Kimi-Researcher as a paradigm shift for agentic AI: a system that makes several decisions on its own to complete a complex task. In his view, it goes beyond producing fluent answers and shows expert-level autonomous reasoning, the kind of complex cognitive work earlier LLMs lacked.

Taken together, these releases show how fast competition in AI is intensifying, and how seriously Chinese companies are now challenging the largest US players on both capability and price.