KEY POINTS

 

Alibaba-backed startup Moonshot unveiled its Kimi K2 model late Friday night — an open source, low-cost large language model with strong coding capabilities.

 

Moonshot claims Kimi K2 outperforms Anthropic’s Claude Opus 4 on two benchmarks and surpasses OpenAI’s coding-focused GPT-4.1 on several industry metrics.

 

Meanwhile, OpenAI CEO Sam Altman announced an indefinite delay of OpenAI’s first open-source model due to safety concerns.

 

BEIJING — The latest Chinese generative AI challenger to OpenAI’s ChatGPT comes with coding power — and at a significantly lower cost.

 

Moonshot, backed by Alibaba, launched its Kimi K2 model late Friday. Like China-based DeepSeek’s disruptive move in January, Kimi K2 combines open source access with affordability — an approach rarely taken by major U.S. tech players apart from Meta and, to some extent, Google.

 

Almost simultaneously, OpenAI CEO Sam Altman announced Saturday that the company would again postpone the release of its first open-source model due to safety concerns. OpenAI did not immediately respond to CNBC’s request for comment on Kimi K2.

 

Kimi K2 specializes in generating computer code — a field where businesses see opportunities to cut costs and automate development. Anthropic also targeted this market with its Claude Opus 4 model released in late May.

 

In its announcement on X and GitHub, Moonshot reported that Kimi K2 outperformed Claude Opus 4 on two benchmarks and delivered better overall results than GPT-4.1, according to multiple industry metrics.

 

“It’s no doubt a globally competitive model, and it’s open sourced,” said Wei Sun, principal AI analyst at Counterpoint, in an email Monday.

 

A Cheaper Alternative

Kimi K2 offers lower token costs, making it appealing for large-scale or cost-sensitive deployments, Sun noted.

 

Unlike ChatGPT or Claude, which require paid subscriptions for their latest models, Kimi K2 is accessible for free through Kimi’s app and browser interface. Moonshot charges just $0.15 per million input tokens and $2.50 per million output tokens — dramatically lower than Claude Opus 4’s $15 and $75, or GPT-4.1’s $2 and $8, respectively.

 

Moonshot states that developers can freely use K2 in commercial products, provided they display “Kimi K2” on the interface if the product has over 100 million monthly active users or generates over $20 million in monthly revenue.

 

Positive Early Reception

Early feedback on K2 from English and Chinese social media has been largely positive, though some users reported occasional hallucinations — a common generative AI flaw.

 

Pietro Schirano, founder of AI design startup MagicPath, said on X: “It’s the first model I feel comfortable using in production since Claude 3.5 Sonnet.”

 

Moonshot previously open-sourced some of its models. Its chatbot gained traction last year as a homegrown alternative to ChatGPT, which remains officially unavailable in China. However, competition has since intensified with offerings from ByteDance, Tencent, and Baidu, which integrated AI into its core search engine.

 

The launch comes amid investor interest in Chinese AI firms challenging U.S. dominance. Despite DeepSeek’s earlier momentum, it has yet to release major upgrades to its R1 and V3 models, while another startup, Manus AI, recently moved its headquarters to Singapore.

 

In the U.S., OpenAI has yet to release GPT-5. Counterpoint’s Sun suggested engineering resources might be focused on GPT-5, delaying the company’s open-source ambitions — especially given the risk of undercutting its proprietary model’s advantage.

 

A Grok 4 Rival

Kimi K2 isn’t Moonshot’s only recent project. Last month, the company launched the Kimi research model, claiming it matched Google’s Gemini Deep Research score of 26.9 and outperformed OpenAI’s version on the “Humanity’s Last Exam” benchmark.

 

Elon Musk’s xAI recently introduced Grok 4, which scored 25.4 on that benchmark — or 44.4 when enhanced by AI tools and web search.

 

“Kimi-Researcher represents a paradigm shift in agentic AI,” said Winston Ma, adjunct professor at NYU School of Law. He described it as demonstrating autonomous expert-level reasoning, moving beyond simply producing fluent responses.

 

Ma, author of The Digital War: How China's Tech Power Shapes the Future of AI, Blockchain and Cyberspace, added: “It’s the kind of complex cognitive work previously missing from large language models.”