Taking advantage of ridiculous token subsidies
Alibaba has taken low-cost tokens to a new level. I've been using Vercel's AI Gateway, where Alibaba is offering Qwen 3.7 Flash for $0.03/million input and $0.13/million output, and we're not talking about a low-tier model or slow token generation.
To be clear, you're sending your data to a Chinese company's cloud, so this isn't something you'd use for private data (always run local models if there's anything private about your data).
- I've been working on open source so privacy is not a constraint.
- I don't vibe code. I use AI as an assistant and put thought into the process. I cannot comment on how Qwen 3.7 Flash would work as a slop cannon.
- Qwen 3.7 Flash is not a frontier model. It follows instructions extremely well but will make mistakes without good instructions.
In the interest of transparency, I didn't know about this model until a week or two ago. My preferred model is Poolside's Laguna S 2.1. That doesn't work though, due to the model routinely being unavailable. And I'm not talking about the free tier. The paid tier is more often than not going to be unavailable.
A couple things surprised me about Qwen's model:
- It's aggressive about caching. I used it heavily for a couple hours yesterday and the total bill came to 6 cents.
- It was able to handle some things that even expensive models struggle with. LLMs in general struggle with vectors holding hundreds of lines with long numbers. Qwen 3.7 Flash sliced right through that problem. I thought maybe it was ignoring accuracy issues, but no, extensive testing was unable to find any errors.
Obviously Alibaba is not making money on this model. I don't care. It's not like I'm going to buy frontier AI to work on open source. I'll take advantage of it to advance open source while the opportunity is there.