HN Explorer

Has Opus 5.5 gotten dumber for anyone else in the last few days?

1 points by geuis ↗ HN
Since it was released Opus 5.5 has solved almost all the problems I've had with previous Anthropic models over the last year. On medium thinking level (default in Claude code) it's been remarkably more succinct and really good at understanding my questions and even anticipating my needs. Has been a markedly huge improvement over previous models and even had me commenting to a colleague the other day that if all advancement were to stop, it would be a useful tool for at least the next 5 years.

But over the last 2-3 days, it's started exhibiting some of the problems previous models presented. It's been making absurd assumptions without asking about my preferences. More definitively, in 3 instances I've had to either correct it on previous decisions that were made but it ignored. 2 times I've had to double check it on some quite obvious high level performance choices I've come to anticipate the model already understanding from prior experience, especially in the same project. At least twice I've had to call it out only get the remedial "you're totally right I should have done X instead".

Also to be clear, I'm not playing around with more advanced expensive models that fall back. This is all experience within a Pro plan using Opus 5.5 medium.

Anyone else been encountering weirdness?

7 comments

[ 0.14 ms ] story [ 15.4 ms ] thread
bbor
Well the biggest disconnect is that I rarely run the orchestrator model at medium -- I just checked the ArtificialAnalysis index, and Opus-5.5-xhigh is about as much better than Opus-5.5-medium than Opus-5.5-medium was than Opus-5-medium (~5 PP, or ~10%). If I'm letting a model plan changes and then orchestrate their implementation without talking through the details first, it's almost always worth it to me to pay for that extra intelligence.

The second biggest difference, though, is that I haven't used Opus 5.5 in a few days cause I ran out of credits on Thursday! So... you might be doing something right, there ;)

One thing you could check is the files that tend to get read into context, especially their built-in "memory" feature, which can get insanely cluttered if not actively maintained. This is why I prefer keeping everything I can to three other abstractions, none of them harness-specific: reports (roughly per-(long-)turn), policies (human readable, very concise, identified with unique slugs, system-wide), and tasks (Plane/Trello/Linear/etc., mentioned in git commits).

For something like "high level performance choices", a policy to always think about performance explicitly before reporting back could honestly do wonders. As you say yourself, they're typically smart enough to know what to do when reminded of your priorities or asked explicitly. The trick is to not prioritize everything, everywhere, all at once!

...so weird talking about how to best use these agents, BTW. Truly shuffling the deck chairs on our Titanic of a career/civilization.

kstenerud
I haven't noticed any difference. In fact, I just set it on a laundry list of 12 items to sort out overnight. It took 10% of my weekly quota to do it, and it probably would have only taken half the tokens if I'd been monitoring it. But it did work while I slept, so I'm not too torn up about it.

There were a few small issues with its work (it's not the greatest at getting layouts right, and one of the features it implemented feels clunky and will need tuning), but the architectural refactors it did worked perfectly and I'm happy with the performance improvements.

I haven't yet checked out the research tasks it did, but usually it does a decent job. I could have just asked it to research and implement, but I find it best to manually verify its conclusions first.

agntflow
I see it the same. Vibe coding is dead. They did something. The intelligence is still there but harder to pull it out. Is less capable from one shot, much less. If project scope is missing well..hard to pull it out, you need much more extra effort
carbongo
Been correcting it for 2 days straight. Something's definitely up
carbongo
Holy fuck! It cannot tell the difference between files. It assumed the file with the exact same name in Trash was the file I was talking about and it nearly put it live...

And this is Opus 5.5! A week earlier it would've been impossible.