The productivity claims are mostly noise in both directions. What I can report is a change in behaviour.

Where it genuinely helps

• Unfamiliar syntax and library surfaces, where the alternative was twenty minutes of documentation.
• The boring middle of a refactor, once the shape is decided.
• Writing the test I was going to skip because setting up the fixture was tedious.
• Exploring an approach I would previously have dismissed as too expensive to try.

Where it does not

Deciding what to build, choosing a boundary, or judging whether a design will survive contact with the second requirement. It is confidently wrong about the things that matter most, and confident wrongness is expensive precisely because it does not look like an error.

The skill that has gone up in value is review. I read more code than I write now, and a good deal of what I read looks plausible and is subtly not what I asked for.