The global AI race has just heated up, and once again, China is leading the way. A new rival has entered the field, Moonshot’s Kimi K2 Thinking, a model so advanced that its creators say it has outperformed OpenAI’s GPT-5 and Anthropic’s Cloud Sonnet 4.5 on several high-profile benchmarks. Yes, the same GPT-5 that has been hailed as the pinnacle of logic and intelligence may have just arrived at its counterpart, and it’s open source.

The AI ​​arms race has seen many dramatic turns, but Moonshot’s latest release may be one of the most surprising. Costing less than $5 million for training and pocket change, compared with the billions of dollars spent on American AI labs, KMK2 Thinking is presenting itself as proof that small players can still make global waves.

What is Moonshot AI’s Kimi K2 model?

Beijing-based AI lab Moonshot, which has been quietly building momentum since its founding, unveiled Km K2 Thinking on Thursday. The company claims the model outperforms GPT-5 and Cloud Sonnet 4.5 on Humanities Last Exam, BrowseComp, and SEAL-0, three key benchmarks that test reasoning, problem-solving, and AI’s ability to dig up hard-to-find information online.

“K2 Thinking is capable of planning, reasoning, executing, and adapting hundreds of steps to tackle some of the most challenging academic and analytical problems,” the company writes on its website. It is powered by a Mixture-of-Experts (MOE) architecture, essentially a brain composed of specialized sub-brains that collaborate to solve complex tasks.

In plain words, it can think ahead, adjust mid-process, verify its own answers, and even use tools like a web browser to refine its reasoning. It’s designed to decompose vague, open-ended problems into concrete, solvable steps that even the best chatbots still struggle with.

Built with a staggering one trillion parameters, the Kimi K2 Thinking is based on the first Kimi K2 model launched in July and is now accessible via Hugging Face for developers to experiment with. In an AI world dominated by pay-walled, proprietary models, this is a statement of intent.

And perhaps the most surprising detail? Moonshot says the whole thing was trained for just $4.6 million, which is reportedly a fraction of the cost to train GPT-5 or Cloud Sonnet 4.5. If those claims hold up under independent verification, it could rewrite the economics of AI innovation.

China is winning the AI ​​race

If the release of DeepSeek’s R1 earlier this year was China’s “Sputnik moment,” the Kimi K2 sighting feels like the country’s first moon landing, pun intended. The arrival of the new models has reignited debate over whether the US can maintain its dominance in artificial intelligence amid increasingly capable Chinese competition.

Adding fuel to that conversation, Nvidia CEO Jensen Huang reportedly told the Financial Times this week that “China is going to win the AI ​​race”. Speaking at the FT Future of AI Summit, Huang cited low energy costs and loose regulations as key advantages for Chinese companies. His comments were later softened in an official statement, where he clarified that “China is nanoseconds behind the US in AI. It is important that the US wins by leading developers around the world.”

Huang has long argued that America’s AI leadership depends on keeping developers dependent on Nvidia’s chips, the same chips that power most of today’s leading models. His efforts to ease US export restrictions on chip sales to China also led to his meeting with President Donald Trump earlier this year, after which Washington reportedly agreed to ease some restrictions.

The rivalry between US and Chinese AI developers has taken a distinctly ideological tone. American AI is often seen as a product of open markets and individual freedom, while Chinese AI is seen as a tool of centralized control, yet ironically the latest model of moonshot is open source, potentially liberalizing the West’s most powerful AI firms.

Whether Kimi K2 Thinking actually dethrones GPT-5 will depend on independent testing, but its release has already generated excitement and concern throughout the industry. After all, if a $5 million Chinese model can match or outperform billion-dollar Western systems, the balance of power in AI could shift more quickly than anyone expected.

As the arms race continues, one thing is clear: it’s not just about smart chatbots anymore. It’s about who defines the next era of intelligence, and whether the future of AI will be shaped in Silicon Valley, or somewhere further east.