Chinese Models Gain Traction with U.S. Firms
Chinese AI models from DeepSeek and Z.ai are gaining significant traction with U.S. companies: the share of tokens routed to Chinese models on OpenRouter has exceeded 30% every week since Feb. 8 and peaked at 46%, up from a 12-month average of 11%, according to CNBC. Z.ai's GLM-5.2 is now being compared by Reuters and The New York Times to Anthropic's Opus 4.8 and OpenAI's models, with several Chinese-developed models appearing on public leaderboards. CNBC reports rising token prices at U.S. labs are a key driver, citing Brookings' Kyle Chan on growing cost sensitivity among enterprise buyers. For practitioners, the shift signals a widening set of cheaper, near-frontier inference options that can lower production costs but add vendor-evaluation and governance overhead.
The practical story here is economic, not just competitive: once a non-frontier model is good enough for a given task, cost gravity pulls production traffic toward it, and Chinese labs are currently winning that trade on price even as reported capability gaps narrow.
What happened
CNBC reports that Chinese-built models from companies including DeepSeek and Z.ai are gaining traction with U.S. firms and are increasingly viewed as competitive with leading American frontier systems. CNBC reports that the share of tokens routed to Chinese models via the developer platform OpenRouter has stayed above 30% weekly since Feb. 8 and reached as high as 46%, versus a 12-month average of 11%. Reuters and The New York Times report that Z.ai's GLM-5.2, released in June, drew comparisons to Anthropic's Opus 4.8 and OpenAI's frontier models, and that several Chinese-developed models now sit on widely watched public leaderboards. Reuters quotes David Sacks: "We now have a Chinese open-weight model that is as good as the currently available models from OpenAI and Anthropic." CNBC quotes Kyle Chan of the Brookings Institution on rising cost sensitivity among U.S. companies, and The New York Times quotes Rehaan Ahmad on the narrowing capability gap.
Technical context
Independent OpenRouter usage tracking corroborates the token-share trend: Chinese-origin labs account for a large and growing share of identified token volume on the platform as of mid-2026, while combined U.S.-origin lab share has fallen back. Reported price differentials of roughly 60-90% versus leading U.S. models are the driver cited across this reporting: when a task does not require the top-tier model, teams are increasingly routing it to the cheapest model that clears the bar.
For practitioners
The reporting points to three operational levers becoming more common: routing noncritical or high-volume tasks to cheaper open-weight models, running multi-model pipelines that reserve premium models for edge cases, and increasing use of caching and batching to exploit lower per-token pricing on commodity workloads. Teams evaluating vendor mix should benchmark end-to-end product metrics (latency, safety, hallucination rate) rather than leaderboard position alone, and should weigh data-governance and export-control questions before routing production traffic to a Chinese-hosted or Chinese-origin model.
What to watch
Independent, standardized benchmark comparisons of GLM-5.2 against Opus 4.8 and current OpenAI models; further OpenRouter token-share data as U.S. labs respond on price; and regulatory or procurement-policy shifts, since data-security concerns remain the main friction point cited for wider enterprise adoption of Chinese-origin models.
Key Points
- 1Chinese models now capture over 30% of weekly OpenRouter token share among U.S. users, peaking near 46%, versus an 11% year average.
- 2Z.ai's GLM-5.2 draws direct comparisons to Anthropic's Opus 4.8 and OpenAI's models as several Chinese models reach public leaderboards.
- 3Teams should benchmark end-to-end product metrics, not leaderboard rank alone, and weigh governance risk before routing traffic to Chinese models.
Scoring Rationale
This is a notable, well-corroborated shift in LLM inference economics because CNBC verifies rising OpenRouter usage while Reuters, The New York Times, and the official GLM-5.2 model card support the capability and cost-performance context. It affects vendor and model-selection decisions broadly, but it remains below major industry-shaking impact until enterprise adoption is shown beyond aggregator traffic.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
