Guys, sorry but I need to vent.
Opus 4.7 is the dumbest model I've ever used from Anthropic. It tries shortcuts that aren't allowed, makes modifications and recommendations completely different from what was requested, and still gives wrong answers or at the very least is too lazy to look up the most recent answer.
And this is with me always asking it to work on the Opus model at maximum effort.
I'm really disappointed. Today I lost an entire day on a session that did everything it wanted and almost nothing of what I asked for.
A session of 159,53 just to produce a rich, properly tailored prompt for my needs on this job. claude-opus-4-7: 4.5k input, 325.8k output, 159.4m cache read, 11.5m cache write (159,53)
And an execution of 243,59 that did everything I didn't ask for and didn't deliver what I did ask for. claude-opus-4-7: 16.7k input, 599.1k output, 413.9m cache read, 3.5m cache write (243,59)
If I had spent 403,12 on API with these sessions — which I'm now reverting to an earlier backup and throwing the day's work away — I would have torn my hair out.
I still have two more weeks left on the 20x plan and I'll keep trying, but I've already discarded any kind of recommendation I was making before they nerfed 4.6 to launch 4.7.
I'm dreading going back to spending money on OpenAI, which is also a company that's made it onto my no-go list.
Does anyone have anything new to recommend, or are we orphaned without an intelligent coding model like Opus was back around mid-February?