Chinese Open Models Took OpenRouter's Entire Top 5, and OpenAI Fell Out of the Top 10

Chinese Open Models Took OpenRouter's Entire Top 5, and OpenAI Fell Out of the Top 10

6 min readJuly 16, 2026

Quick verdict

A screenshot of OpenRouter's monthly model ranking is making the rounds because it looks lopsided: the top five slots by token usage are all Chinese open-weight models, seven of the top ten are Chinese, and neither OpenAI nor Google appears anywhere in the top ten. The Financial Times reports the same trend from the enterprise side, with companies moving to Chinese open weights to cut costs. One caveat matters up front: this is OpenRouter's own routing traffic, not global AI usage. But the direction is real, and the reason is boring in the best way. It is cheaper.

What the ranking actually shows

The dashboard that kicked this off ranks models by monthly token share on OpenRouter. The top five, in order, were DeepSeek V4 Flash, MiMo-V2.5, MiniMax M3, Hy3 preview, and DeepSeek V4 Pro. Chinese models held seven of the top ten. OpenAI and Google, the two names most people would expect near the top, were absent from the list entirely.

Before reading too much into that, keep the source straight. OpenRouter is a routing and prototyping layer. People use it to test many models through one API before deciding whether to self-host or keep paying a provider directly. So the ranking measures what developers route through OpenRouter, not what the whole market runs in production. A lot of ChatGPT and Gemini usage never touches OpenRouter at all. The chart is a signal about where cost-sensitive builders are pointing their traffic, not a scoreboard of who is winning AI.

With that caveat in place, the signal is still worth something. The people who route through OpenRouter are exactly the ones watching cost per token, and they are voting with their volume for open Chinese models.

Why the switch is happening

The most quoted line from the discussion sums it up: it is hard to compare benchmarks, but easy to compare bills. Models like DeepSeek V4 Flash and MiMo-V2.5 are cheap enough through hosted inference that, for many teams, they beat both the closed frontier APIs and the cost of self-hosting once you count electricity and hardware. One structural factor came up repeatedly: electricity in China runs well under half of US prices, which flows straight into inference costs at scale.

There is also a trust angle that has nothing to do with price. Several developers said they distrust closed Western providers because of pricing changes and model churn. A model you depend on can get more expensive, get quietly swapped for a new version, or disappear. Open weights remove that risk. Once you have the file, no vendor can take it away or reprice it out from under you. We saw the sharp end of that when a top closed model went dark under export rules, which is part of why open weights keep reading as a continuity plan, not just a cost play.

The enterprise side, in one story

The Financial Times ran a piece titled "Companies Turn to Chinese Open Weight Models to Cut Costs," which lands on the same point from the boardroom rather than the terminal. And a widely shared thread from a Fortune 500 "AI First" org showed what the pullback looks like in practice. After a failed pilot that tried to have agents reverse-engineer a legacy app into a formal spec, and day-to-day code generation that kept producing unreliable output, the company stopped its AI demos, cut broad Claude access, and asked teams to limit usage or move to older and cheaper models.

The trigger that made the cost visible was pricing. Commenters pointed at GitHub Copilot's shift to usage-based billing as the moment large organizations could finally see what heavy agent workflows actually cost. Flat per-seat pricing had hidden the consumption. Metered billing exposed it, and once it was visible, the cheaper model became the obvious default for everything that did not truly need the frontier. This is the same routing logic that cut enterprise AI bills nearly in half when companies applied it deliberately rather than in a panic.

The policy backdrop

The competitive pressure has reached Washington. According to a widely cited report, the Trump administration and industry groups discussed streamlining US open-model releases of equal or lesser capability to the leading Chinese open models, as a direct response to how competitive the Chinese local ecosystem has become. The debate splits along a predictable line: US labs are wary that releasing strong open weights could cannibalize their own paid API and SaaS revenue.

On the other side of the Pacific, Zhipu founder Tang Jie defended open-sourcing frontier AI outright, arguing that model security comes from transparency and broad oversight rather than restricted access. Zhipu has released GLM-5.2 under an open license for download and commercial use, part of the steady stream of capable open checkpoints we tracked in our open-source AI models guide. Whether that openness is principle or strategy, the effect on the cost curve is the same.

Why it matters if you are not running a data center

The enterprise version of this story is a company routing cheap requests to a cheap model and saving a fortune. The individual version is the same idea at a smaller invoice. If you pay separate flat fees for ChatGPT Plus, Claude Pro, and Gemini Advanced, you are locked into three premium models and paying full price whether the task is hard or trivial. That is the opposite of what the cost-conscious developers on OpenRouter are doing. The personal equivalent of routing is one account that reaches every model, so you can send the easy questions to the cheap option and save the flagship for the work that needs it. That is the whole case for an AI aggregator over a stack of subscriptions, and the same logic driving both the OpenRouter chart and the FT headline.

Video: the rise of cheap Chinese AI models

A rundown of how Chinese open-weight models went from underdogs to the default cost-cutting choice.

FAQ

Does this mean Chinese models are better than GPT or Gemini?

No. The ranking measures token volume on a cost-sensitive routing layer, not raw capability. The frontier closed models still win the hardest tasks. What the chart shows is that for the large majority of everyday requests, a cheap open model is good enough, and price decides the rest.

Why are OpenAI and Google missing from the top ten?

Mostly because a lot of their usage never runs through OpenRouter. People reach ChatGPT and Gemini through their own apps and APIs. OpenRouter skews toward developers shopping for the cheapest capable model, which is where open Chinese models are strongest.

Are these open models safe to run for real work?

For a large share of tasks, yes, which is why the FT found companies moving to them to cut costs. The common setup keeps open models for volume and a premium model in reserve for the hardest problems. See our open-source AI models guide for where they stand.

How do I get this cost advantage as an individual?

Stop paying separate flat fees locked to one model each. Use one account that reaches every model and pick the cheapest one that can do the job. That is routing at a personal scale, and it is the core argument for an AI aggregator.

Sources

Further reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.

Start free on Admix

Related articles