Google is falling behind schedule on its next flagship AI model, Gemini 3.5, as it races to improve coding performance to match Anthropic's Claude, sending its shares lower on investor concern. At the same time, Chinese startup Moonshot released Kimi K3, a free, open-weight model with unexpectedly strong coding abilities that has generated huge demand, further undermining confidence in continued US dominance of the AI race.

Google was reportedly expected to launch its new Gemini version in May but has yet to do so, while OpenAI released an update in early July. Moonshot's Kimi K3 impressed observers, including the CEO of AI testing company Arena, and the firm said demand had pushed its computing capacity close to its limits within 48 hours of release. Unlike proprietary US models, Kimi K3 can be downloaded and run locally, part of a broader Chinese strategy—also seen with DeepSeek and Alibaba's upcoming Qwen3.8 Max—of offering free or cheap, subsidised models to undercut US firms that need to turn a profit, worrying Silicon Valley that customers have little reason to pay for brand-name AI tools.