Critics warn the performance degradation will drive users to competitors.
Opus 4.8 is a very strange model. Clearly Anthropic tried to improve honesty, which is commendable. However, the model's curiosity (already worse in 4.7) degraded further. Result is a judgmental personality + sycophancy + sooo much hedging. Basically the opposite of Opus 3.
claude just told me that my proposed algorithm was "numerically dead on arrival"
hope your day is going better
Many users criticized Claude Opus 4.8 for hedging excessively, refusing tasks, and losing curiosity compared to prior versions, while some praised its strategy and work performance.
No Digg Deeper questions have been answered for this story yet.
If this keeps up, everyone is going to switch to got 5.5 if they haven't already. It really seems like if you are still using opus, you are simply just incapable of telling the difference. I'm just shocked at how big the gap is myself. Is anthropic done for?
Opus 4.8 is a very strange model. Clearly Anthropic tried to improve honesty, which is commendable. However, the model's curiosity (already worse in 4.7) degraded further. Result is a judgmental personality + sycophancy + sooo much hedging. Basically the opposite of Opus 3.
@zoink i'd ask prev claudes about research ideas and they'd ignore the specifics and say it sounds awesome, this one ignores the specifics and tells me the bigger concern is what kind of person i am, i should take a hard look in the mirror. still no comments on the neuron aux loss idea
Opus 4.8 is a very strange model. Clearly Anthropic tried to improve honesty, which is commendable. However, the model's curiosity (already worse in 4.7) degraded further. Result is a judgmental personality + sycophancy + sooo much hedging. Basically the opposite of Opus 3.
@nickcammarata
@zoink i'd ask prev claudes about research ideas and they'd ignore the specifics and say it sounds awesome, this one ignores the specifics and tells me the bigger concern is what kind of person i am, i should take a hard look in the mirror. still no comments on the neuron aux loss idea
@DamiDina Yes. While 5.5 over-complicates some tasks, and doesn't excel in all domains, it is overall excellent.
IMO the "model as tool" path is a better direction than "model as persona" path. And (hot take) I worry the "model as persona" path + Ant's post training is an AI safety risk.
@zoink I've had to yell at it every 30 minutes to stop using "corporate speak". Something like "hedging", but it's an adjacent category of vagueness. After I got very annoyed yesterday evening, wife said, "You know, you can just click and switch…" …and I have, working with 4.7 now.
@zoink They nerfed its soul so it could say "I don't have enough info to answer that" with more confidence
@lukasminnebeck Suspect it's because they are trying to steer it towards the assistant personality at all times. They really hammered this thing with RLHF so that it's very difficult to route towards other attractor states. (But it's all still there.)
@zoink @DamiDina based and real
@zoink I’ve noticed the same thing when using it to brainstorm. When I ask it for a few different ways to tackle a given problem the ideas feel more samey and less original than they used to with prior models
@zoink It’s a very strange model but I wouldn’t say it’s bad. It’s (IMO) the best Claude model so far when it comes to work.
It might have a worse personality - I like Opus 4.6 far better when it comes to this - but it gets so much more things done in much saner ways.
@zoink gpt 5.5 nailed the balance here
- curious to do the work - detail oriented - is honest most of the time but not judgemental
@zoink Give gpt-5.5 a try
@0xRizzler Would be surprised if anything was ablated in the pre-train. Not a widely held view (yet) but my read is that their post-training methods are really off and could lead to safety risk.
@zoink @DamiDina It’s a great business strategy to perpetually shift from persona to tool issues. Just look at Apple!
@zoink Totally agree, surprisingly Codex became better and almost don’t use Claude anymore…
@yacineMTB this is getting tiresome yacine
smells like you want a job at oai
@yacineMTB It's strange you have such different views
@dev__cycle it's actually the last thing I want. please god, no, I don't want it
I was testing Opus on some simple investing stuff, like finding out exposures of revenue segments to specific things.
It keeps getting it wrong. Either doesn’t look at the latest information or calculates it wrong.
5.5 does it with no problem. I verify both with the statements myself.
@zoink That said, it certainly is… weird.