Chinese Models Now Run Nearly Half of US API Traffic

For two years the assumption inside American AI was that the frontier belonged to a short list of US labs, and everyone else was playing catch-up. A CNBC investigation published July 7, built on routing data from the model-aggregation platform OpenRouter, quietly demolished that assumption. Chinese-origin models now account for as much as 46% of the enterprise tokens flowing through one of the most-used developer gateways in the West — and they have held at least 30% of weekly volume every single week since February 8, 2026.

A year earlier, that number was 4.5%. Averaged over the prior twelve months it was 11%. The line didn't drift; it broke.

The number that should stop you

OpenRouter sits between developers and dozens of model providers, routing each API call to whichever backend the developer picks. That makes its aggregate token flow one of the cleanest available proxies for what US engineering teams are actually running in production — not what they say in surveys, but where the requests go. And the requests are going east.

In the measured peak week, combined Chinese models hit 46.4% of routed tokens against 35.7% for all US models put together. The single largest vendor on the platform wasn't OpenAI or Anthropic. It was DeepSeek, at 17.6% of routed tokens — roughly 5.13 trillion per week. Alibaba's Qwen followed at 13.9%, about 2.77 trillion tokens weekly. Anthropic, the biggest American provider by this measure, sat at 14.8%. The Chinese share first overtook the US share in the week of February 9–15, when Chinese models pushed 4.12 trillion tokens through the pipe.

This is not a story about consumer chatbots or benchmark leaderboards. It's a story about the invisible plumbing — the API calls that power other companies' features — and in that plumbing, the balance of power has already tipped.

Price is doing all the work

The explanation is almost boringly economic. Open-weight Chinese models are, in OpenRouter's own framing, "consistently 60% to 90% cheaper" than the leading US offerings, per the platform's Justin Summerville. At the extremes the gap is a chasm: DeepSeek V4 Flash lists at $0.14 per million input tokens, while OpenAI's GPT-5.5 runs $5.00 for the same million. Across the board, Chinese models come in anywhere from 4x to 100x cheaper than their American counterparts.

For a solo developer, that difference is a rounding error. For a company running billions of tokens a month through a product feature, it's the entire margin. When a model is good enough for summarization, classification, extraction, or routine code generation — the unglamorous 80% of real production traffic — engineering teams do what engineering teams always do: they route the task to the cheapest option that clears the quality bar. And the recent wave out of China keeps clearing it.

That's the mechanism nobody priced in. The frontier labs assumed capability was the moat. But most production workloads don't need the frontier. They need reliability at a price that survives contact with a CFO. Once open-weight models got close enough on quality, the decision stopped being about which model is smartest and became about which model is cheapest-that-works — and that is a race the high-margin American incumbents are structurally built to lose.

What the token flow doesn't capture

There are caveats worth stating plainly. OpenRouter skews toward cost-sensitive developers and indie builders; it is not a census of the Fortune 500, many of whom still route through Azure, Bedrock, or direct enterprise contracts where compliance, data-governance, and procurement inertia weigh heavier than per-token price. A regulated bank or a defense contractor is not going to pipe customer data into a DeepSeek endpoint because it saves fourteen cents a million tokens.

Governance is the live wire here. The same week's news cycle underscored how quickly the ground can move: US export-control directives briefly forced the suspension of frontier American models before they were restored days later — a reminder that model access is now a policy variable, not just a product one. For anything touching sensitive data, "where does this model run and who can see the tokens" is no longer a paranoid question; it's a procurement checkbox. That friction is real, and it caps how far the cost advantage can carry Chinese models into the most regulated corners of the market.

But it does not un-happen the 46%. The developers who move first and fastest — startups, tooling companies, the builders wiring up the next generation of agents — are precisely the ones OpenRouter captures, and they have already voted with their traffic.

The uncomfortable read

The strategic picture this paints is the one American labs least wanted to see. The open-weight ecosystem, largely written off in 2024 as a fast-follower that would never quite catch the closed frontier, has instead become the default substrate for cost-conscious production. China's labs didn't win by building a smarter model than the best in the West. They won by making a good enough model radically cheaper and giving it away as open weights, then letting the market's relentless cost-optimization do the rest.

For US frontier labs, the message in OpenRouter's data is sharp: capability leadership at the top does not automatically translate into volume leadership in the middle, where the tokens — and eventually the revenue — actually live. The moat everyone was defending was the frontier. The battle that mattered was the floor. And on the floor, the cheapest thing that works is winning the trade.