Is it just me, or has Opus suddenly turned dumb as rocks with 0 attention span over the last half day or so? Noticeably worse output across multiple agents on very different projects.
Agreed. Yet, it munches tokens like there's not tomorrow. I don't get anything done with Claude in the past 2-3 weeks - just eating up its weekly quota within a day.
It is dumb. None of my threads were closed with shipped features. It keeps going into endless review cycles essentially burning all my quota within a couple of days. And I'm not talking about complex a features an OpenAPI-based integration and some enhancement of pytest traces building maps of touched files. Both these are in their second week of token burn!
There was an article on HN a month or so about benchmarking Opus 5.5 every day and recording / publishing the results. I recall this being a github repo, but my searches turned this up instead: https://www.bridgebench.ai/nerf-bench/claude-opus-5-5
So it seems like ... it is not getting dumber. I find myself using Fable a lot more, though. Opus 5.5 was never as fun as Fable, but it's pretty good.
I’ll ask the same question as when it came out: if you don’t trust the provider how are you trust them not to look for and cheat on repeat benchmarks like this.
I think something has changed, maybe about two weeks ago, because there’s this fixation from the model where if you ever need to make a correction, it simply becomes disinterested with your instructions.
Is it just me, or has Opus suddenly turned dumb as rocks
What I've seen is that if I let Opus 5.5 run for too many turns (not sure how to quantify that) it starts exhibiting signs of human dementia. It will forget it told me to run something and will then accuse me of misconfiguring something that it literally gave me a script to run that configures something. It also forgets the instructions in my project and has to go back and read them and apologizes for forgetting the instructions. Between this increasing context rot and the excessive shock collars applied to defensive security coding it's rare that I can even use Opus 5.5. For the past few weeks I have had to use Opus 4.8 and even that now has a sub-set of the shock collars applied [cyber] which forces me to start a new session. This has led me to researching the Abliterated 70B/120B (uncensored) models and required hardware.
So dumb as rocks? I am not seeing that, but the context rot is getting worse for me personally. I don't know if it makes a difference, I am on the pro 20x plan. Maybe people on enterprise plans do not see this which I assume is most of the people here using Claude.
21 comments
[ 0.25 ms ] story [ 25.9 ms ] threadMy best guess was also that in one or two days they will drop a new model that will be so much more better than the nerfed 5.5.
So it seems like ... it is not getting dumber. I find myself using Fable a lot more, though. Opus 5.5 was never as fun as Fable, but it's pretty good.
https://github.com/ninjahawk/livenerf
I’ll ask the same question as when it came out: if you don’t trust the provider how are you trust them not to look for and cheat on repeat benchmarks like this.
What I've seen is that if I let Opus 5.5 run for too many turns (not sure how to quantify that) it starts exhibiting signs of human dementia. It will forget it told me to run something and will then accuse me of misconfiguring something that it literally gave me a script to run that configures something. It also forgets the instructions in my project and has to go back and read them and apologizes for forgetting the instructions. Between this increasing context rot and the excessive shock collars applied to defensive security coding it's rare that I can even use Opus 5.5. For the past few weeks I have had to use Opus 4.8 and even that now has a sub-set of the shock collars applied [cyber] which forces me to start a new session. This has led me to researching the Abliterated 70B/120B (uncensored) models and required hardware.
So dumb as rocks? I am not seeing that, but the context rot is getting worse for me personally. I don't know if it makes a difference, I am on the pro 20x plan. Maybe people on enterprise plans do not see this which I assume is most of the people here using Claude.