US Firms Turn to Cheaper Chinese AI Models

US technology companies are increasingly moving AI workloads away from premium models made by OpenAI and Anthropic and towards cheaper Chinese alternatives, according to reporting by The Washington Post on 14 July. The shift has accelerated through June and July as businesses confront rapidly rising bills for the AI systems now embedded in customer support, coding, internal tools and workplace automation.

Chinese model families including Alibaba’s Qwen, Z.ai’s GLM, Moonshot AI’s Kimi and DeepSeek are benefiting from the change. The attraction is simple: they can offer competitive performance at dramatically lower prices, particularly where companies are able to use open-weight models or host them through specialist providers rather than rely solely on the most expensive closed systems.

San Francisco startup Lindy.ai offers one of the clearest examples. The company makes AI assistants that handle tasks including email and calendars, and had previously relied heavily on Anthropic’s top-tier models. But chief executive Flo Crivello said Anthropic had become the company’s largest expense, ahead of payroll for more than two dozen staff and ahead of rent. Last month, Lindy moved 100% of its traffic to DeepSeek-V4.

“By far, our No. 1 expense was Anthropic.”

Crivello said the Chinese model was “just 10x cheaper” and that the change had saved Lindy millions of dollars. “So it was a very, very simple business decision,” he said.

Cost is reshaping the AI market

The pricing gap is substantial. Justin Summerville, who works on data and analytics at OpenRouter, said open-source Chinese models can be 60% to 90% cheaper than leading Anthropic and OpenAI offerings. Artificial Analysis benchmarking, as reported by Vested Finance, put the cost of a standardised workload at $4,811 on Anthropic’s Claude and $3,357 on OpenAI’s ChatGPT, compared with $1,071 for DeepSeek, $948 for Kimi and $544 for Zhipu’s GLM.

That makes Claude nearly nine times more expensive than the cheapest Chinese alternative in that comparison. For an individual developer, the difference can be even easier to understand: an hour-long coding session that costs about $10 on Claude cost less than 50 cents on DeepSeek, according to the brief. DeepSeek V3.2 is priced at $0.28 per million input tokens, versus roughly $10 per million for GPT-5.2, a difference of around 35 times.

The movement is not limited to smaller firms. Airbnb and Siemens are experimenting with moving daily operations towards Chinese AI companies including Alibaba and DeepSeek. A UBS report said larger global enterprises are becoming more interested in open-weight models as token spending becomes a significant concern. It cited an unnamed major global bank that had started hosting Alibaba’s Qwen models to control costs while balancing their use with premium systems such as Claude.

Coinbase chief executive Brian Armstrong has also said his company is experimenting with making open-weight options the default, including GLM 5.2 and Moonshot’s Kimi 2.7, while allowing engineers to select models depending on the task.

Usage is already climbing

The adoption data suggests this is more than a handful of cost-conscious experiments. OpenRouter reported that Chinese models accounted for more than 30% of tokens used by US companies on its platform every week since 8 February, reaching as high as 46%. That compares with a 12-month average of 11%, and just 4.5% in the first half of 2025.

On Vercel’s AI Gateway, daily token volume for Zhipu’s GLM-5.2 rose 50-fold from mid-June. DeepSeek V4 Flash became the platform’s largest individual model by volume, taking more than 20% of traffic. Open-weight models represented 29% of token volume on Vercel’s platform, nearly three times their share in April.

Alibaba’s Qwen has built the biggest model ecosystem on Hugging Face, with more than 113,000 derivative models. A Hugging Face study published on 16 March found Chinese open-source models accounted for 41% of downloads.

Kimi K3 raises the stakes

The latest high-profile release arrived on 16 July, when Beijing-based Moonshot AI launched Kimi K3. The 2.8-trillion-parameter mixture-of-experts model has native visual understanding and a one-million-token context window. Moonshot says its full weights are due by 27 July, when it will be possible to self-host the model or run it through open-model providers.

On independent testing, Kimi K3 ranked fourth among frontier models, behind Claude Fable 5 and GPT-5.6 Sol but ahead of Claude Opus 4.8. It opened at number one on the Arena.ai WebDev leaderboard with a score of 1,679, above Claude Fable 5’s 1,631 and GPT-5.6 Sol’s 1,618.

K3 is not a bargain-basement model: its $3 per million input tokens and $15 per million output tokens make it Moonshot’s most expensive release yet, and place it around Claude Sonnet pricing. Still, its arrival underlines how quickly Chinese labs are closing the gap at the top end, while older models continue to provide the biggest savings.

Why it matters

For ordinary users, model choice is likely to show up indirectly. If app makers and workplace software firms can cut their AI running costs, more features may become viable without steep subscription increases or strict usage limits. The trade-off is that businesses will need to judge performance, reliability, hosting arrangements and data handling carefully rather than choosing a model on price alone.

US companies still lead in the race for the most capable frontier models, with experts saying Chinese systems remain six to 12 months behind on capabilities. But China’s push into open-weight software, shaped partly by US restrictions on advanced Nvidia hardware, has created a powerful alternative for firms that do not need the absolute best model for every job. The next test will be whether premium American AI providers respond with lower prices, clearer value or still greater capability.