A Mozilla report shows frontier AI models from US companies like Anthropic offer only a 4.4-month performance advantage over Chinese open-source alternatives like Moonshot's Kimi K3, while costing 5 times more. Most organizations should default to cheaper open models for routine work, reserving expensive closed models for specialized tasks requiring expert professional work, intensive retrieval, or long context processing.
Anthropic accused Chinese AI companies DeepSeek and Moonshot of secretly routing user requests to Claude models and presenting the responses as their own, exposing Chinese users' data to the American company. The allegations escalate tensions following US government claims that six Chinese firms stole US AI model outputs, with Beijing rejecting such accusations as suppression tactics.
Anthropic released a threat intelligence report documenting how Claude was misused for cyberattacks, influence operations, surveillance, and weapons development, with a notable case involving a Chinese AI company (Moonshot) serving Claude instead of its own model while collecting user exchanges for training data.
The article disputes Anthropic's claims that Moonshot and DeepSeek route requests to Claude models, arguing instead that Chinese AI gateway services (transit stations) exploit discounted subscriptions and resell access to multiple models while profiting by routing cheaper alternatives and selling user data to Chinese labs.
Chinese AI labs DeepSeek and Moonshot secretly relayed customer prompts to Anthropic's Claude through fraudulent accounts, exposing sensitive data including military and government information. The labs also saved exchanges to train their own models, with Moonshot forwarding nearly 300,000 requests over ten days.