Every major model release is accompanied by a wave of excitement and a wave of skepticism, and both are equally useless if you don’t test it on your own tasks. I ran the new version through what I do every day - code, reviews, drafts - and share what has actually changed, and not what is written in the announcement.
Where things got noticeably better
The first thing that catches your eye is that the model holds the task context longer and is less likely to “slip” in the middle of a long analysis. Previously, on large-scale tasks, it was necessary to remind people about the initial conditions; Now this is needed significantly less often.
Secondly, working with code has become more careful in the smallest details: fewer fictitious methods, more often an honest “I’m not sure, check this.” For vibe-coding, where you quickly assemble and immediately race, this saves exactly those minutes that usually go into catching hallucinations.
Where people are still needed
A strong model does not cancel editing. In public texts, she still tends to use streamlined language that looks sleek but doesn't assert anything. The final pass is still up to you - especially where you need a position, not a retelling.
What does this change in practice?
Nothing is automated by itself from the fact of a new release. Something else is changing: the bar of “good enough to give to work” is moving, and part of the routine, which was previously easier to do by hand, is now cheaper to entrust to models with verification. This is the real effect - not magic, but a shift in the boundary.