China Took Over Open-Source AI
61% of tokens on OpenRouter now route to Chinese models. Four of the top five open-weight models are Chinese. The takeover is measurable — and it runs on price.
For two years the story of frontier AI was an American one. But underneath it, a different race was quietly decided. By the middle of 2026, when a developer anywhere in the world reached for an open-weight model, the odds were they reached for a Chinese one.
The Takeover Is Real
OpenRouter now routes ~61% of tokens to Chinese models. Four of the top five most-used open-weight models are Chinese-made. Llama dropped off the list. The numbers are measurable, not rhetorical.
The download crossover happened earlier: by late 2025, Chinese model downloads (17.1%) had already overtaken US downloads (15.86%). Then came May 2026 — four frontier-adjacent open models shipped in a 12-day window. DeepSeek V4. Qwen 3.5. Kimi K2. GLM-5. The velocity was deliberate.
But "Took Over Open" Needs an Asterisk
Here's the catch: the best open Chinese model still trails the top proprietary US models by ~6–9 points on the benchmarks that matter. OpenAI, Anthropic, and Anthropic's partners still own the frontier. They're not losing that.
What they are losing is the floor beneath it. Menlo's recent analysis found open-source is only ~11% of enterprise production API usage — down from 19% two years ago. In the Fortune 500, the story is still American closed models. "Dominant in tokens," as one researcher put it, "marginal in Fortune 500 production."
The volume is in the tokens. But the profit and control? Still in the premium US labs.
Why China Went Open
It wasn't accidental. Four labs, four strategies:
| Lab | Model family | Where it presses hardest |
|---|---|---|
| DeepSeek | DeepSeek V4 | Coding and raw price-performance |
| Alibaba | Qwen 3.5 | Broadest ecosystem and reasoning |
| Moonshot | Kimi K2 | Agentic tool use and long tasks |
| Z.ai | GLM-5 | Top-tier open-weight coding |
The logic is old: commoditize the complement. If the model layer becomes a fungible commodity, the value shifts upstream (to the infrastructure and the data) and downstream (to the applications and services built on models). It's the same playbook that destroyed the margin on browsers, operating systems, and databases.
There's an irony embedded in the strategy. US export controls limited Chinese compute. That constraint forced efficiency. The efficiency produced a cost advantage. The cost advantage let them out-compete on price even with less hardware. The controls backfired — they accidentally subsidized the very thing they were trying to contain.
The Price War and What It Does
A DeepSeek API call costs ~$0.01 for input tokens, ~$0.03 for output. OpenAI's GPT-4o is 10–30x more expensive, depending on volume. Qwen undercuts OpenAI by a factor of 5–10. The price gap is structural — lower training cost, smaller models, efficient inference, and a different margin philosophy.
When the market sees a 20x price difference for "good enough," it compresses the price the market will pay for anything not clearly frontier-best. It's pressure on OpenAI's and Anthropic's economics. Not a threat to their frontier models. A threat to the margin that funds them.
The Bottom Line
The US labs are not losing the frontier. They are losing the floor beneath it, and the floor is where most of the volume lives.
The Chinese strategy is patient. Set the default across Asia. Make open-weight the expected baseline in developer communities. Push the US labs' closed premium model from "default" to "luxury." In three years, when an enterprise evaluates "do we pay 20x for the frontier, or do we use open?" the answer starts to change.
Open-source is how you win the long game when you can't win the short one.
Are Chinese models better than American ones?+
On the frontier, no. Best Chinese open models trail top US proprietary models by ~6–9 points. But on price and sufficiency, Chinese models win decisively.
Which is the best Chinese open-source model in 2026?+
No single winner. DeepSeek V4 dominates coding. Qwen 3.5 offers broadest reasoning. Kimi K2 leads agentic tool use. GLM-5 is top-tier for code. All four cost 5–30x less than comparable US models.
Why did Chinese labs open-source when US labs did not?+
Commoditize-the-complement strategy. If the model layer becomes fungible, value shifts upstream to infrastructure and data, downstream to applications. Export controls forced efficiency, creating a cost advantage.
Is it safe for enterprises to use Chinese models?+
Depends on use case and regulatory environment. Open-source adoption is ~11% of enterprise production API usage. Fortune 500 still overwhelmingly uses closed US models. Startups and developers lead adoption.