Okay, I realized a couple of things this weekend, and honestly I’m still kind of amazed.
I canceled my ChatGPT and Claude subscriptions and went all in on Chinese models and their cloud stack. I also used the DeepSeek API platform directly for the first time (not via OpenRouter).
What I found was pretty wild.
DeepSeek may be a bit old at this point, I’m still waiting for V4, but their setup is so good that I spent just $2 on R1 and used it for hours. I still haven’t even burned through that $2.
In the terminal, it gets surprisingly close to what I want from Claude Sonnet 4.6, and it’s much, much faster.
If I tried the same amount of compute with a Western provider, I’d probably be down $20 to $30 already, and there’s a good chance I still wouldn’t have finished the task because of the compute bottlenecks they’re dealing with right now.
That’s the part that really got me.
Asia is catching up fast. The whole “they’re 6 months behind” line just does not look true to me anymore. And honestly, a lot of these benchmarks are not measuring what actually matters in real use.