Chinese AI Models Now Dominate Developer Traffic

DeepSeek's wordmark. The Chinese AI lab is the single largest vendor on OpenRouter by token volume. Image: RoadMaster19, CC BY 4.0, via Wikimedia Commons.

Update, Sept. 29

MiniMax quietly shipped a coding-only model, Startup Fortune reported September 27 — aimed at entry-level coding workloads, following the same playbook described below: ship often, price low, publish weights. It is further evidence the release cadence is accelerating, not stabilizing.

In February, Chinese AI models handled between 6% and 13% of the token traffic on OpenRouter, one of the largest marketplaces developers use to route requests across AI models. By the week of September 14, that share was 57% to 67%.

In seven months, the balance of power in applied AI flipped — and almost nobody outside the industry noticed until the number got big enough to alarm people in Washington.

CNBC reported Friday that Chinese models are rapidly increasing adoption by global companies on the back of better performance and lower prices. The data, shared with CNBC by the platforms themselves, points to a structural shift, not a blip.

The numbers, measured in actual workloads

On OpenRouter, Chinese-made systems accounted for 57% to 67% of all tokens processed during the week of September 14, versus 6% to 13% in February. On Vercel, the share of Chinese models rose to 55% of tokens in August, up from 11% in January. Two platforms, the same direction — measured in real developer workloads, not surveys or sentiment.

The leaders are names American developers now route to by default. DeepSeek is the single largest vendor on OpenRouter by token volume, routing roughly 5.13 trillion tokens a week and holding 17.6% of the platform's total traffic. Alibaba's Qwen follows at 13.9%, with 2.77 trillion tokens weekly. Moonshot AI's Kimi K3 has muscled into the same tier, with the company claiming it beats every model on the market except the newest releases from OpenAI and Anthropic.

Agentic coding is the workload driving much of the traffic shift — Chinese open models can now credibly handle it at a fraction of the price. File photo: Martin Vorel, CC BY-SA 4.0, via Wikimedia Commons.

The blunt explanation is price. As of June, DeepSeek's V4 Flash model cost $0.14 per million input tokens; OpenAI's GPT-5.5 cost $5.00 for the same unit. Open Chinese models are running 60% to 90% cheaper than the leading US alternatives, according to figures OpenRouter shared with CNBC. For the agentic coding and customer-service workloads that now eat most enterprise AI budgets, that gap is too large to ignore.

Frontier US models still pull in more total spending — large enterprises paying for cutting-edge reasoning haven't abandoned OpenAI or Anthropic. But the volume of everyday, high-frequency traffic, the stuff that actually costs a company money at scale, has moved to Chinese models fast. A real chunk of US token demand is now behaving like a commodity, routed to whichever open model is good enough for the job.

Why the models are good enough now

This is not just a price story. It is a capability story that happened while the West was arguing about safety pauses.

Peter Walker, head of OpenRouter Insights, told CNBC that Chinese open-source models released this year "can credibly perform in advanced agentic use cases, especially in regards to coding," and are "incredibly cost-effective compared to most models from American labs." Harpreet Arora, head of agent infrastructure at Vercel, said Chinese models have improved enough to handle more work at lower cost — and once a model clears the quality thresholds needed for business, the price gap becomes the deciding factor. For more complex tasks, he added, companies still want cutting-edge US models.

The release cadence tells the same story. Alibaba shipped Qwen3.8-Max on August 3, a 2.4-trillion-parameter flagship priced at $2 per million input tokens, then followed on August 12 with the first open-weight version of a Max-tier Qwen model. Zhipu's GLM-5.3 landed August 14 with a million-token context window. DeepSeek and Kimi run the same playbook: ship often, price low, publish weights. The pattern — near-constant releases designed to keep developers evaluating Chinese models instead of settling on one Western default — is working.

None of these are closed products you pay a subscription for. They're open-weight models anyone can download, run on their own servers, and build a business on top of. And that is exactly what is making Washington nervous.

Washington is sending letters

Two US House committees are investigating the impact of wider adoption of Chinese models, and the alarm is specific. Daniel Remler, a senior fellow in the technology and national security program at the Center for a New American Security, told CNBC that Chinese AI represents "real economic and security risks for the United States" and warned that integrating Chinese models could pull countries into China's technology sphere of influence.

The United States has restricted Chinese AI companies from buying cutting-edge chips through export controls, but Washington is concerned they may narrow the gap by remotely accessing Nvidia chips through overseas data centers or by using distillation methods that mimic existing models — the same distillation techniques the US flagged in a September 9 security advisory.

Earlier this month, the US accused six Chinese AI firms of copying models and bypassing chip controls. The Stanford HAI AI Index 2026 found the US-China AI capability gap had narrowed from as much as 31.6 percentage points in 2023 to just 2.7 percentage points by March 2026. Chinese models now account for approximately 41% of downloads on Hugging Face over the past year, with cumulative global downloads of Chinese open-source AI models exceeding 10 billion.

Chinese AI models FAQ: why developers switched

How big is the price gap? As of June, DeepSeek's V4 Flash cost $0.14 per million input tokens against $5.00 for OpenAI's GPT-5.5 — and OpenRouter's figures shared with CNBC put open Chinese models 60% to 90% cheaper than the leading US alternatives. For the agentic coding and customer-service workloads that now eat most enterprise AI budgets, that gap is too large to ignore.

Which Chinese models lead? DeepSeek is the single largest vendor on OpenRouter by token volume, at 17.6% of the platform's total traffic (roughly 5.13 trillion tokens a week). Alibaba's Qwen follows at 13.9% with 2.77 trillion tokens weekly, and Moonshot AI's Kimi K3 has muscled into the same tier.

What is Washington actually doing? Two US House committees are investigating the impact of wider Chinese-model adoption. The US has restricted Chinese AI companies from buying cutting-edge chips through export controls, flagged distillation techniques in a September 9 security advisory, and earlier this month accused six Chinese AI firms of copying models and bypassing chip controls. The Center for a New American Security's Daniel Remler warned Chinese AI poses "real economic and security risks for the United States."

Does this mean China leads in AI? On developer traffic volume and price, China is winning the everyday workload. On capability, top US frontier models still lead: the Stanford HAI AI Index 2026 found the US–China gap had narrowed to just 2.7 percentage points by March 2026 — but it still favors the United States.

The open-weights trap: why tariffs can't fix this

Here is the uncomfortable part for American AI strategy. Closed-weight models — the kind produced by Anthropic, OpenAI, and Google — keep their parameters proprietary; any country or enterprise using them is operationally dependent on the producing company's API, pricing, and policy decisions. Open-weight models release their parameters publicly, letting any government or company download, run, and modify them on sovereign hardware without any ongoing relationship with the original developer.

China has bet heavily on open weights — partly by necessity, after US chip export controls limited access to cutting-edge training hardware, and partly by strategy. The result is a form of technological influence that no tariff or export control can claw back: once the weights are downloaded, the model belongs to whoever runs it.

Trump administration officials, for their part, are treating AI chiefly as a national-security competition. The recent Trump–Xi summit established a dedicated channel for managing potential AI incidents, and Treasury Secretary Scott Bessent has said US–China AI talks would explicitly cover "both open and closed weight models." The framing acknowledges the reality: the fight is no longer just over who trains the best model, but over whose models the world actually runs.

One caveat is worth keeping. The OpenRouter and Vercel figures are platform samples with different time windows and geographic coverage — they are strong distribution evidence for DeepSeek, Alibaba, and peers, not proof of majority use across all AI or of leadership on frontier benchmarks, where top US models still lead. Token volume is winning the headline. Margin, capability, and influence are what will decide the decade.

But the margin question cuts both ways. If the everyday workload of the AI economy increasingly runs on Chinese open models at a tenth of the price, the US labs face a squeeze that no amount of frontier-model supremacy fully answers: the best models in the world don't pay the bills if the volume goes somewhere else.

Read more: US Accuses Six Chinese AI Firms of Copying Models and Bypassing Chip Controls

Sources: CNBC (reporting via Startup Fortune, Digital Today, TechSambad, Sept. 25–28, 2026); Stanford HAI AI Index 2026; Tech Times; CNAS; OpenRouter; Vercel.