I'm not using any marketing; I'm sharing my experience. At work, I recently had to build an interface for JNI in Lua. I had to do a lot of the fiddly bits myself, but most of the interface the LLM generated without a problem, just being fed my existing Lua interfaces and the JNI docs. I had to do some clean up and optimization, but it definitely saved time. I then fed it the JTOpen Java docs and had it use the interface to make a wrapper around JTOpen for communicating with IBM i message queues (which was the original goal of the project), and it did enough of a good job to definitely be worth the expense (notably, Claude did nearly perfect at generating JNI method signatures that it inferred from the Javadocs; as a counterpoint, it did generate interfaces for a couple methods that didn't actually exist). These things are really really good at generating interface glue, generating good documentation, generating tests, and limited refactoring.

They won't do the thinking for you. You still have to know what you want out of it, you have to know your endpoint, and you have to be able to review and understand every line out of it, but they do save time. You might have tried to use them in situations where they aren't appropriate. They are text prediction programs, but they're extremely powerful text prediction programs. If you accept their unreliability and give them the right patterns and context to generate from, they can write a thousand lines of code in one go that look identical to what you would have written, and it takes 30 seconds to generate and a few minutes to review and test, compared to the half an hour at least that it would have taken otherwise. It can easily yield 10 times efficiency gain in places where it's really appropriate.

That said, it's not perfect by any means. To be really what I want, I need the same quality significantly faster, cheaper, and local. If my company wasn't paying for it, I certainly wouldn't pay to use this thing myself. This costs between $0.05 and $1 per request, which gets expensive fast, and the much cheaper models are less capable enough to save me so little time as to not be worth the bother really.