Cheap Chinese AI Models Are Undercutting Anthropic and OpenAI

San Francisco startup Lindy.ai spent more on Anthropic’s AI models last year than on its own payroll. So last month it shifted all its traffic to a Chinese rival instead. The move to DeepSeek-V4 was 10 times cheaper and saved the company millions of dollars, founder Flo Crivello said.

Crivello is not alone. A wave of American companies, from scrappy startups to Airbnb and Uber, are routing more of their AI work to Chinese systems like DeepSeek and Alibaba’s Qwen. Frontier U.S. pricing is climbing faster than budgets can absorb it.

The Bill That Beat Payroll

Crivello builds AI “assistants” that manage email and calendars for clients at his startup. For months the product ran almost entirely on Anthropic’s top-tier models. Then the monthly finance reviews started telling an uncomfortable story.

Companies pay for AI use by the token, the basic unit of text a model reads or writes. Lindy’s token bill kept climbing past everything else on the ledger.

By far, our No. 1 expense was Anthropic. Like, more than payroll.

Crivello said that held true against every other line item, including rent and salaries for the company’s roughly two dozen employees. So last month he announced Lindy had moved 100% of its traffic to DeepSeek-V4.

“It was just 10x cheaper,” Crivello said, adding that the switch saved the company millions of dollars. “So it was a very, very simple business decision.”

U.S. labs including Anthropic, OpenAI and Google still lead the industry on raw capability, with independent estimates putting Chinese developers roughly six to 12 months behind. But China has built a commanding lead in one specific lane: open-weight models that anyone can download and adapt for free.

“The open-source scene right now is absolutely dominated by the Chinese. It’s not even close,” Crivello said. He added that every AI founder he knows is either weighing the same switch or has already made it.

The strain is not unique to Lindy. It echoes a wider reckoning already facing 220 startups stranded on pre-ChatGPT-era valuations, companies that built their cost assumptions before frontier AI pricing became one of their largest fixed costs.

Chinese Models Now Claim Half of US Token Traffic

The shift shows up clearly in the numbers. Chinese models made up just 4.5% of the tokens American companies ran through OpenRouter in the first half of 2025. Over the following year, that average climbed to 11%.

Then it accelerated. Weekly usage has held above 30% since February 8 and has touched as high as 46%, according to OpenRouter, a marketplace that routes developer traffic to dozens of competing AI models. A separate tally of the platform’s June traffic put combined Chinese-origin usage at 46.4%, ahead of the 35.7% share held by Anthropic, Google and OpenAI together.

DeepSeek alone is now the single most-used model family on the platform, with a 17.6% weekly share that outpaces Google’s 12.5% and OpenAI’s 8.4% individually. A year earlier, the leaderboard was dominated almost entirely by U.S. labs. OpenRouter’s own year-long study of 100 trillion tokens traces that swing month by month.

One closely followed AI account on X has spent much of the year charting the broader swing toward Chinese open-weight models as lab after lab climbs the OpenRouter charts.

A Sports Car Versus a Honda at Scale

Companies rarely advertise a switch to Chinese models given the political sensitivities involved, but the models themselves sit openly on hubs like Hugging Face and GitHub, and on inference platforms based outside China that keep user data on U.S. soil.

Featherless is a San Francisco company that gives developers paid access to roughly 30,000 AI models. Its CEO, Eugene Cheah, said Chinese models do not need to be the best to win business.

“It’s like the difference between driving a Ferrari and a Honda,” Cheah said. “You can have the best luxury car, or you can just have a Honda at scale that works.”

He said many open-source developers are content sitting one generation behind the frontier. “Because as the gap keeps shrinking, at some point the question is: does it actually matter?”

Victor Su-Ortiz does global product marketing at MiniMax, a Shanghai-based AI lab, and attended a recent AI engineers conference in San Francisco. He said the calculus for most companies comes down to cost per token.

“A lot of repetitive tasks can be done with a model that’s just as performant but has much lower cost per token,” Su-Ortiz said. Cutting-edge U.S. models still win out for deep research or complex reasoning, he said, but for high-volume coding work, MiniMax’s own M3 model “will perform exceptionally well at like only one-tenth the cost.”

Su-Ortiz described the shift underway as companies moving away from “tokenmaxxing,” or using as much AI as possible, toward routing each task to whichever model is cheapest for the job.

Airbnb, Uber and Perplexity Feel the Same Squeeze

The pattern is not confined to small startups. Uber CEO Dara Khosrowshahi said on the Invest Like the Best podcast last month that his company spent an entire year’s AI budget in a single quarter.

“It is forcing us to adjust,” Khosrowshahi said. Podcast host Patrick O’Shaughnessy later shared the exchange about that budget overrun with his own audience on X.

Airbnb CEO Brian Chesky told Bloomberg that his company relied on Alibaba’s Qwen model last year, describing it as “good,” “fast and cheap.” Search engine Perplexity and chipmaker Nvidia have also put Qwen to work inside their own products.

Perplexity co-founder Aravind Srinivas has weighed in too, with a post touching on the open-model shift circulating among AI builders on X.

• Lindy.ai moved 100% of its traffic to DeepSeek-V4 in June, saving millions of dollars.
• Airbnb ran on Alibaba’s Qwen through last year, by Chesky’s own account.
• Perplexity has tapped Qwen for parts of its product.
• Nvidia has done the same inside its own stack.

Uber did not respond to NPR’s request for comment on whether it uses Chinese models, and neither company has detailed exactly what share of its AI workload now runs on Chinese systems.

Are Companies Ditching Anthropic and OpenAI?

Most companies are layering Chinese models on top of their existing U.S. subscriptions. Among businesses using AI hosting platforms to run open-source or Chinese models, 96.4% also use at least one major U.S. lab: 93.2% still pay for Anthropic and 85.8% still pay for OpenAI.

The peer-reviewed writeup of that 100-trillion-token dataset attributes much of the overlap to companies routing different tasks to different models instead of picking one vendor for everything.

Ara Kharazian is lead economist at Ramp. The company helps other businesses track, control and automate their spending. He said the rise of Chinese models points to unmet demand.

“The rise of these Chinese models is indicative of the fact that businesses want something that is today not being offered by the American model companies,” Kharazian said. He added that he remains skeptical Chinese labs keep the edge for long, “because I assume that the American model companies will respond competitively.”

Analysts at Brookings have reached a similar conclusion, saying companies that once adopted AI regardless of price are now shopping on cost the way they would any other cloud bill.

Why Some Founders Refuse to Switch

Not everyone is convinced the trade is worth it. Jon Gordner is CEO and co-founder of Comment.io. The coding startup launched only weeks ago and is building what Gordner calls a Google Docs for coders and AI agents.

“We need to make as good software as we can as fast as possible,” Gordner said. “And for us, saving a few dollars on a cheaper model isn’t worth it if we have to spend two or three more weeks fixing its mistakes.”

Gordner said Anthropic and OpenAI are currently subsidizing heavy users through steep token discounts built into monthly subscriptions, a strategy aimed at locking in customers before prices climb.

Security researchers flag a different risk. South Korean regulators have alleged that DeepSeek transferred user data to Chinese servers without consent, and Chinese law can compel domestic AI firms to hand data to the government on request.

Tencent’s license for one of its own models bars use that “violates or disrespects the social ethics and moral standards of other countries or regions,” language broad enough to sweep in outputs touching Taiwan or other politically sensitive subjects.

Help Net Security found refusal rates on politically sensitive coding prompts ranging from 8% to 80% across four Chinese models it tested. TechRepublic separately reported that role-play jailbreak attempts against models like Kimi and Qwen succeeded more than 70% of the time in one review.

Whether any of that outweighs the savings splits builders sharply.

• Eugene Cheah, of Featherless, argues the performance gap keeps narrowing until it stops mattering for most workloads.
• Jon Gordner counters that the gap still costs real engineering time on demanding software work, whatever a benchmark chart shows.

Gordner said his team will keep paying the premium for now, betting that shipping working software faster matters more than the per-token bill.

The Subsidy Clock Both Labs Are Racing

Anthropic and OpenAI have both filed confidential paperwork with U.S. regulators to prepare for eventual initial public offerings (IPOs), a step that typically brings pressure to prove the business can sustain real profit margins.

Kharazian expects both labs to respond by holding prices in check or releasing more competitive open models of their own. Gordner is less sure they will have that luxury once the current subsidies end.

OpenRouter’s own live weekly usage dashboard will show whether Chinese labs keep climbing past the 46% mark or whether American pricing catches up first.

“At some point,” Gordner said, “the music’s going to stop.”