I see a lot of speculation that this is somehow due to the recently-announced "watermarking", but that seems really unlikely to me, because the watermarking is about nudging per-token output probabilities or the sampler in a detectable way. That's not going to change the structure of every output to honest caveat this and verbing abstract nouns that. For the examples people are showing, and I've seen plenty of gibberish myself, you don't get there by perturbing probabilities from any sane starting point.

it started at Opus 4.7, way before watermarking