So you’ve burned through your AI tokens again, and now you’re sitting there waiting for them to unlock. That whole dance - the rate limits, the waiting, the paying more to wait less - is a big part of why everyone is drifting toward Chinese models. Everyone except you, of course.
The numbers are moving fast
Let me show you a chart. This one comes from OpenRouter, an aggregator that gives you access to both Western and Asian models through their APIs - so you pay purely per usage.
Last year, non-American models accounted for about 10% of usage on OpenRouter. This year that figure has climbed to 40%.
And then there’s the price. Compared to the top of the range - Anthropic’s models - the Chinese models in the same tier cost a quarter of the price. Half of half.
The usual objection is privacy: “I trust the Americans more.” Do you, though? At this point you’ve already handed all your data to the Americans, so you might as well give it to the Chinese too and let them fight it out. What do you care?
What this actually means in practice
The Chinese open models - whether you use them through an API or run them locally - enable a strategy the closed vendors don’t.
- Start on the cloud. You’re up and running immediately, and it costs very little.
- Move in-house later if you want. Because these models are open, you can eventually bring them home with a local setup. That comes with its own headaches, but the option exists.
That’s the key point: closed vendors like OpenAI and Anthropic simply don’t let you make this move. With open models you keep the exit door open.
The catch: performance
People say the high-end Chinese models are on par with the Western ones. That’s not really true - especially compared to Anthropic.
Here’s a concrete example. The subtitles you see in my videos are generated through a set of skills, and inside those skills are scripts that do a whole chain of things. Setting that up with Claude was immediate. It barely made any mistakes, and it worked the moment I stood it up.
Doing the exact same thing with the high-end Chinese models - I tried both a top-tier one and GLM - I had to rewrite all the skills and all the scripts, make them far more explicit, far easier to follow, so the underlying language model could actually get the work done.
So what do you choose?
It comes down to two postures:
- Pay more, stay lazy. Sit back on the couch, belly out, and let the premium model handle everything.
- Pay less, roll up your sleeves. Use the Chinese models, save a fortune, but do the extra work to make them behave.
Send this video to a friend who’s going through token withdrawal. Bella regà.