Chinese AI models moved deeper into U.S. use as developers and companies adopted lower-cost systems, with Kimi downloads and OpenRouter rankings showing rapid growth.

Lower prices and stronger performance are drawing developers and companies to Kimi, GLM and other systems despite security and provenance disputes.

Mozilla technology chief Raffi Krikorian said he switched many daily tasks to Moonshot's Kimi K3 and had also used Z.ai's GLM-5.2.

Coinbase and other U.S. users cited by AP described Chinese models as a way to reduce inference costs.

The five most-used models on OpenRouter during the preceding month were Chinese, according to platform data cited by AP.

Sensor Tower estimated that Kimi downloads exceeded 930,000 during the week after K3's release, up 200 percent week over week.

U.S. Kimi downloads were estimated at about 86,000 during that period, a 387 percent increase.

Moonshot temporarily suspended new subscriptions after demand pushed available capacity close to its limit.

AI services charge for input and output tokens, and agent workloads can multiply inference calls, increasing the importance of unit cost.

U.S. officials and companies have alleged that some Chinese labs improperly distilled capabilities from U.S. systems; China rejects those accusations.

Independent evaluators quoted by AP said Chinese models still trail U.S. frontier systems across the full range of capabilities despite strong performance on many tasks.

The evidence supports a bounded conclusion: Chinese AI models moved deeper into U.S. use as developers and companies adopted lower-cost systems, with Kimi downloads and OpenRouter rankings showing rapid growth. It does not resolve the remaining uncertainty because Download and routing data do not measure confidential enterprise use, reliability, data handling or performance in every workload. The factual horizon is therefore defined by Independent safety and capability evaluations and U.S. procurement, export-control or data-security rules affecting use.

The current reporting limit is specific: Download and routing data do not measure confidential enterprise use, reliability, data handling or performance in every workload.

The next verifiable developments are Independent safety and capability evaluations and U.S. procurement, export-control or data-security rules affecting use.