I see both types of claims a lot on HN, and find both credible.
On one extreme, a solo dev could prompt a frontier SOTA agent with not much more than "Build Uber", give it the necessary hardware/software/permissions to build itself a useful feedback loop and act autonomously, put it in goal mode, and let it cook, and as long as the tokens kept flowing they'd eventually end up with something reasonable in actual orders of magnitude less time than it would have taken that dev to build the same thing — especially if you're counting calendar days rather than hours worked, and extra especially if they wouldn't have had the luxury of investing full-time hours into that particular thing. At the opposite extreme, BigCo full-time SWE #5000 tasked with an extremely specific ticket for a critical revenue-generating brownfield system might experience even "20%" as liberal, and correctly observe that writing code was never the bottleneck for their job. And then there's the massive spectrum of circumstances in between those two ends.
None of that means that the gains aren't real, or that anyone in particular is wildly exaggerating in either direction. It just means the gains aren't evenly distributed.