China’s AI models are shrinking US lead
Daniel Howley· Technology Editor
Sat, July 18, 2026 at 12:25 AM GMT+5:30
3 min read
China’s Moonshot AI on Friday revealed that its newest AI model, Kimi K3, is nearly as powerful as some of the US’s leading frontier models, including Anthropic’s (ANTH.PVT) Fable 5 and OpenAI’s (OPAI.PVT) GPT-5.6 Sol
The reveal has heightened concerns that China’s AI companies are further eroding the US lead in the AI race
According to Moonshot, while Kimi K3’s overall performance still falls behind Fable 5 and GPT-5.6 Sol, it still offers frontier-level capabilities on a variety of tests. The announcement comes after OpenAI debuted Sol on July 9. Anthropic revealed Fable 5 on June 9
Both the US and China see the AI race as key to their respective economic and national security efforts. Companies in both countries continuously release new models that outpace those of their predecessors, and any slowdown could give their international rivals the chance to take the lead
The battle for AI supremacy also highlights a key difference between the approaches the countries’ biggest AI firms are taking toward the technology
While companies like OpenAI, Anthropic, and Google (GOOG, GOOGL) focus on releasing proprietary closed-pSeek, and others instead offer their models as open-weight software, which users can download and customize for free
Companies that provide open-weight models often charge customers to run the software on their servers, or, in some instances, if they intend to use it for large commercial purposes
Open-weight models are also cheaper to run than closed- looking to reduce their AI spending, which has become a major cost factor for companies that regularly use AI services
Kimi 3, in particular, is shocking because, in addition to its reported capabilities, it’s also much less expensive to use than both GPT-5.6 and Fable 5. AI companies charge customers by millions of tokens used. A token is a unit of measurement for AI models that is equal to a small word or a portion of a larger word
Moonshot AI says it charges $3 per million input tokens, or prompts you send to a model, and $15 per million output tokens, which are the model’s response
Anthropic, meanwhile, charges $10 per million input tokens for Fable 5 and $50 per million output tokens. OpenAI charges $10 per million input tokens and $45 per million output tokens for GPT-5.6 Sol
Moonshot AI, however, isn’t without controversy. In February, Anthropic accused the company, along with DeepSeek and MiniMax, of an “industrial-scale” campaign to fraudulently “distill” its AI model
