If you are not Experimenting with Open weight. You should be. Especially if it's 18 times cheaper.

I work in Big tech and we have unlimited token counts for all the frontier models etc. this is a gift but also a bit of a vice. I am super detached from what it actually costs to build software with AI tools. But a couple weeks ago I was helping a friend out with some Open source library stuff (ArchUnitGo) and I decided to experiment with opencode. This got me thinking even more. Can a low parameter openweight model do what a frontier does given a solid harness? so I decided to build the library twice. Once with deep seek v4 flash and one with Opus 5.5. And the results are bananas. Same library built and here is the price difference. Opus 5.5: $346.70 V4-flash: $19.86 That's an 18 times price diff. Opus was obviously better overall and took about 40% less tokens to do it but at the end of the day who cares? It's cheap and it works if you use the simpler model. I think we need to really consider what model rightsizing means for the future.