I mean eventually open weight models will be good enough to do the few things AI is actually good at, and then there’s another reason to not pay astronomical prices.

But the benchmarks are semi useless. Performance at new problems in your precise use case matter and that's only something one can find out you using it. This can differ wildly from what benchmarks suggest. AI can be an ability enhancer when for example someone who has no idea hos to code suddenly manages to create crappy code but that manages to do a job where for example success can be easily verified. But if the result should ve good or even reliable ... well....