jokes aside, this is plausible, to the extent that there are feature(s) that detects high-quality output, which there should be to some extent—otherwise i don’t see how something like reflection could work.still trying it myself but so far it feels impossibly good. this is at the level of what i’d expect gpt-4.5 to be, and it’s so goddamn fast. and this is still sonnet! we’re not even at opus 3.5 yet. if anthropic does intend to release opus 3.5 in the next few weeks and the difference between it and sonnet is as large as 3.0 opus was from sonnet, i’m probably going to lose my mind.