← sources

Does anyone else notice that Opus 5.5 is VERY GOOD at the first implementation of a task (one-shotting), but it doesn't follow through as good afterwards?

source: reddit (reddit:ClaudeCode)published 2026-10-08 20:06ingested 2026-10-09 02:39

retained text

I just started noticing this weird pattern where if I give Opus a brand new task with a fresh concept and not updating or adding onto something, it performs a LOT better with "starting from scratch" rather than improving upon something?

I'm working on a game right now and I've just had this happen regularly, it's almost like the models built for oneshotting or something, it's very well able to implement the concept, but fleshing it out isn't able to keep that momentum going and perform as well as it's first try

From my experience it's almost better off using Max reasoning on the first attempt and letting it churn, then lower usage to xHigh/High aftewards on reiterations

document fetched 2026-10-09 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1x119vt/does_anyone_else_notice_that_opus_55_is_very_good