July 8, 2026. US companies are quietly moving serious workloads onto Chinese AI models, and the driver is the line item every operator has been watching all year: cost. CNBC reported Tuesday that open-source Chinese models now run 60 to 90 percent cheaper than leading OpenAI and Anthropic models, and adoption data suggests the discount is winning. By Wednesday, CNBC reported US lawmakers have begun probing how widely these models are used inside American companies.
The numbers behind the shift
- The price gap is an order of magnitude. Leading Chinese models charge as little as 18 cents per million tokens against roughly 4 dollars per million for top US frontier models, per CNBC's reporting.
- Usage share has jumped. On routing platform OpenRouter, the share of tokens US companies send to Chinese models has stayed above 30 percent every week since February 8, peaking near 46 percent, versus an average of about 11 percent across the previous 12 months.
- New releases are landing fast. Z.ai's GLM 5.2, released in June, saw the fastest adoption of any model Vercel has tracked in 2026, with daily token volume growing roughly 27x and customer count roughly 80x in its first full week.
- Whole companies are switching. Automation startup Lindy moved 100 percent of its traffic from Claude models to DeepSeek in June, with its CEO projecting millions of dollars in savings within months.
- Washington noticed. The July 8 lawmaker probe adds a compliance dimension: expect disclosure questions about which models touch which data.
What it means for operators
This is the same discipline we flagged when the AI stack went usage-metered and when Fable 5 moved to usage credits this week: route by task. When a job does not need the best model, send it to the cheapest one that passes your own eval. Bulk classification, extraction, summarization and first-draft generation are exactly where a 90 percent discount compounds, while customer-facing, high-liability and agentic work often still justifies frontier pricing like Sonnet 5's.
The risk column is real but manageable. Data residency and client procurement concerns mostly attach to calling China-hosted APIs, not to the open weights themselves, which US infrastructure providers can host domestically, a distinction we covered when Kimi K2.7 went open weights. The practical checklist: keep an abstraction layer so models swap without rewrites, log which model processed which data class, run Chinese open weights only on US-hosted infrastructure for client work, and get ahead of the disclosure question before your enterprise customers or their lawyers ask it. Building automations that stay model-agnostic while costs swing like this is core to what our AI automation team and AI engineers do for clients.
Frequently Asked Questions
It depends on how you run them. Calling a China-hosted API routes your data through infrastructure subject to Chinese jurisdiction, which many compliance teams will not accept. Running the open weights on US-hosted infrastructure such as major cloud or inference providers keeps data domestic while capturing most of the cost advantage. Log which model touches which data class either way.
Most are released as open weights, so hosting is competitive and priced near compute cost rather than as a premium API. CNBC's reporting puts leading Chinese models as low as 18 cents per million tokens versus roughly 4 dollars for top US frontier models, a 60 to 90 percent discount that widens further with efficient serving.
No. The winning pattern is routing by task: send bulk extraction, classification and drafting to the cheapest model that passes your own quality evals, and keep frontier models where accuracy, safety and liability justify the price. Companies that switched wholesale, like Lindy's move to DeepSeek, did so after testing their specific workloads.
CNBC reported on July 8 that lawmakers are probing the growing use of Chinese models inside US companies, focused on how widely they are deployed and what data they process. For operators the near-term impact is disclosure risk: expect customers, investors and regulators to start asking which models power your products.