Chinese AI Model Breaks into the Elite, Rivals Claude and ChatGPT

Chinese artificial intelligence company Moonshot unveiled Friday its Kimi K3 model, a 2.8-trillion-parameter system that the company claims is the world's largest open-weight AI model and delivers performance close to that of market-leading models.

On its official account, Kimi asserts that, based on its published benchmarks, the model ranks alongside cutting-edge systems such as OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5. The full model weights will be published by July 27.

Moonshot specifies that Kimi K3's training was optimized for complex, long-duration tasks, with sustained programming capability and minimal human supervision. The firm demonstrated a 3D open-world video game built with the model, as well as capabilities for building a GPU programming system from scratch, chip design, and video editing.

An open-weight model means that the company publishes or distributes the files containing the model's learned numbers—its "weights"—so that individuals or companies can download, install, and run them on their own servers, without being obligated to depend on a single platform. This does not necessarily equate to "open source": the program that makes it run may remain closed, and licenses and conditions may limit who can use it and for what purposes.

Moonshot described it as its new flagship model for long-range coding—sustained programming tasks—and for executing complete processes involving research, writing, and internal verification. It also targets use with large contexts, capable of managing information-heavy workflows.

The collaborative evaluation platform Arena AI reported that the model advanced in its web interface programming ranking. "Kimi-K3 is now number one in Frontend Code Arena, surpassing Claude Fable 5," the platform wrote, noting a 17-position jump from Kimi-k2.6, its predecessor model. Arena AI added that Kimi-K3 led six of seven domains in that category and placed second in video games, behind Fable 5.

The company's competitive pricing is also generating discussion in the global market. For enterprises, AI model services are charged per token—fragments of text processed by the system. One million tokens roughly equate to hundreds of thousands of words. Kimi K3 charges $3 per million input tokens, $0.30 per million cached input tokens—reused at a lower cost—and $15 per million output tokens; GPT-5.6 Sol charges enterprises nearly double: $5, $0.50, and $30, respectively.

Following the launch, users posted demonstrations of the model in programming, animation, video game development, and website design.

According to Western media reports, Moonshot AI stated that its system still remains "slightly" behind the highest-performing proprietary models in the United States—Claude Fable 5 and GPT-5.6 Sol—but that the gap is narrowing rapidly.

David Sacks, former head of AI for the Trump administration, wrote: "This is concerning. For the first time, a Chinese model, Kimi K3, has ranked number one in Frontend Code Arena and is scoring at or near frontier level on other benchmarks." He added that, in the meantime, the United States "is getting tangled up" with data center restrictions, state regulations, and new agencies to pre-approve frontier models. "This is how you lose the AI race," he concluded.