Chinese companies are staging quite a comeback in the AI race.
Driving the news: Moonshot, one of China’s top AI startups, is set to release a new model that can reportedly exceed the performance of those at frontier U.S. labs, including Anthropic’s Claude Opus 4.8 — its second-most powerful publicly available model.
According to the Financial Times, Moonshot’s Kimi K3 will be the largest AI model ever created by a Chinese company and will be free to download (though it still charges per-token).
Catch-up: Chinese models still aren’t on par with top-tier systems like Anthropic’s Fable, but the performance gap is shrinking as the price gap widens. The best open-weight models are 10–60 times cheaper than those at top labs, which have raised their token prices over the past year to offset massive computing costs.
Why it matters: Companies blowing through their AI budgets on tokens from Anthropic and OpenAI are realizing that they don’t necessarily need the best-performing (and most expensive) models for every task.
Airbnb, DoorDash, and Siemens are a few of the major companies that have recently adopted Chinese AI tools. DoorDash even tiered its AI model usage based on the work, using Moonshot for lower-level tasks and Anthropic’s Fable for high-level tasks.
Zoom out: OpenAI and Anthropic have accused Chinese startups of mass distillation — the practice of training smaller models on the outputs of top-tier systems — which allows them to replicate their performance without paying for the costly computing resources.
What’s next: The U.S. government (which has already restricted public access to top models) could ban American companies from using Chinese models over national security concerns.—LA
