Lindy didn't switch models to make a point. The AI startup moved 100% of its traffic from Anthropic's Claude to DeepSeek, and told CNBC the switch will save it millions.
It's not an isolated call. Data published by OpenRouter, the API platform that lets developers route requests across more than 400 models from over 60 providers, shows Chinese models overtaking their American rivals in raw usage since the start of 2026. By June, Chinese models were processing roughly 18 trillion tokens a week through the platform's top nine models, against about 5.5 trillion for US models, according to Financial Times reporting cited by Dealroom. In January, the US still led.
The Reversal, By the Numbers
CNBC reports that the share of tokens US companies route to Chinese models through OpenRouter has held above 30% every week since February 8, and has peaked as high as 46%. Over the twelve months before that, the average was just 11%.
Zhipu AI's GLM-5.2, DeepSeek, and Moonshot AI's Kimi are doing most of the pulling. On Vercel's infrastructure alone, daily token volume for GLM-5.2 grew 27-fold in its first week of availability, according to data the company shared with CNBC.
Price Is Doing Nearly All the Work
Vercel's Harpreet Arora, the company's head of agentic infrastructure, put it plainly to CNBC: "Price is doing the work here. When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough."
The gap he's describing is not subtle.
| Model | Provider | Output price (per million tokens) |
|---|---|---|
| Claude Opus | Anthropic | $75.00 |
| DeepSeek V3.2 | DeepSeek | $0.42 |
Chinese open-weight models broadly run 60% to 90% cheaper than leading US systems, per CNBC's reporting. That gap used to come with a real capability tax. It mostly doesn't anymore: GLM-5.2 landed within a percentage point of Claude Opus 4.8 on at least one agentic benchmark at roughly a fifth of the cost, and Brookings researchers estimate Chinese labs now trail top US rivals by six to nine months, not years.
Why Practitioners Are Making the Switch Now
The pressure is coming from both directions at once: American model prices climbing while usage explodes. CNBC reported that Uber burned through its entire annual AI budget just four months into 2026, as Claude Code usage grew faster than the company could tie it to any specific product it shipped. OpenRouter COO Chris Clark told CNBC that Chinese models are "disproportionately represented in agent workflows run by US companies."
Vercel CEO Guillermo Rauch described the same shift from his side of the infrastructure in a July 6 interview with TechCrunch. More than a trillion tokens now flow through Vercel's own AI gateway every day, and he says customers are no longer picking a single lab and building everything on top of it. "We're seeing a lot of growth of Gemini, even though it's not on the news as much, because people are optimizing for production now," Rauch said. "You also bring in open models, so Deepseek and GLM-5.2 are taking off. The data doesn't lie."
The Risk Nobody's Pricing In
The cheapest model available isn't always the most reliable one, and the sector has already had a preview of what that costs. In February, GLM-5 briefly led the OpenRouter charts before demand outstripped Zhipu AI's own compute capacity. Service outages followed, then a public apology, then a price increase. The company's stock fell 22% in a single day.
That precedent is shaping how experts frame the risk today. Hugging Face's Yacine Jernite warned that practitioners risk getting stuck between two bad options: expensive US proprietary models with pricing that can move overnight, or Chinese models as the only affordable way to own their own stack, each carrying different data-security and geopolitical tradeoffs. In our view, that framing understates how immediate the decision already is for teams in finance, healthcare, or government contracting, where the choice of model vendor is being made this quarter, not debated as a future hypothetical.
The Bottom Line
Model capability converged faster than most forecasts expected, and once quality gaps stopped being the deciding factor, procurement teams did what procurement teams do: they started buying on price. Goldman Sachs analysts cited by the Financial Times expect AI agents to drive a 24-fold increase in token consumption by 2030, which means whoever wins the price war on inference is set up to win a much bigger fight over the next several years.
The uncomfortable part for US labs is that this isn't a benchmark loss they can patch with the next release. It's a procurement decision already being made, one engineering team at a time, and the token data shows which way it's currently breaking.
Related coverage: LDS has tracked this shift as it built, from DeepSeek V4 matching frontier benchmarks at one-twentieth the API cost of Claude Opus to GLM-5.2 beating GPT-5.5 on coding benchmarks for one-sixth the price to DeepSeek's own efficiency gains that made its models 85% faster without new hardware.
Sources
- Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge — CNBC, July 7, 2026
- Chinese AI models seize up to 46% of US developer use — Resultsense, citing CNBC, July 7, 2026
- Chinese AI models overtake US peers in token consumption, OpenRouter data shows — Dealroom, citing Financial Times, July 2026
- Vercel CEO Guillermo Rauch on the fight to split off models from agents — TechCrunch, July 6, 2026
- Chinese AI Models Overtake US Rivals in Global Token Consumption — Trending Topics, March 26, 2026