Is it just me, or has Claude Opus gotten worse recently?

12 points by de6u99er 33 days ago | 22 comments

Since the recent updates, I have the feeling that Claude Opus is becoming dumber on complex tasks.

Things that used to work cleanly in a single prompt now fail completely. In coding workflows, it constantly ignores mandatory CLAUDE.md project rules, makes unsolicited edits to unrelated files, and breaks working code. It starts arguing based on stale comments, fails to update documentation when code changes, and edits based on blind guesswork instead of actually verifying the codebase first.

I even caught myself using Fable for tasks that I always used Opus for in the past, just to get decent results.

It feels like Anthropic is shifting their technical problems onto paying users. Either they are using strict safety filters that silently fall back to cheaper models without telling us, or they are downgrading the compute under heavy load to save money.

The result is the same kind of shrinkflation: we pay the same subscription price, but we either burn way more Opus tokens on endless retries, or we are forced to spend money on more expensive Fable tokens to get the quality we used to have.

Has anyone here successfully moved away from Claude for complex engineering workflows? What are you replacing it with, and how are you handling the transition?

Disclaimer: As a native German speaker, I used Gemini to clean up and polish the English text for this post.

yulaow 33 days ago |

I noticed a few weeks ago it started being very bad at explaining things (even things itself was doing) and started committing absurd errors (like reading a test of 5 lines and not noticing there was an explicit mock created in one of those, then saying that the test was failing while it was not)

I fear this is just the classic "nerf the model just before we release a new version of it"

ghoul2 32 days ago | |

This seems a worsening that seems to have started at Opus 4.8. When I (was forced to) switched from 4.7 to 4.7, I noticed that 4.8 had a very strong bias against viewing tool output or reading files. It ALWAYS wanted to 'grep' things. So for example, it would run the test suite and grep for 'PASS', thus completely missing the 4 tests that FAILed. Its bias is so strong, that it does this despite explicit instructions to the contrary, and continues to do so even after having noticed its own mistake, in the same (claude code) session.

For editing files, it again seems to prefer writing small python scripts, rather than read the file and rewrite it (even small files, like notes). This again causes it to constantly miss duplications, inconsistencies, etc.

And from 4.8 to 5, it doubled down on this. At this point not moderately-long-horizon task is getting done consistently, due to its pigeon-hole view of the workdir.

Anybody else notice this?

yulaow 32 days ago | | |

oh yes, exactly the same things are happening to me

spottedmarley 33 days ago |

I know Opus is finished with the simple task I asked for when it starts outputting a 400-page lesson in verbosity (an opus?) complete with comparison tables, bullet-pointed lists of vaguely worded assertions, several self-blunder reports and a list of things it wants to mention but didn't touch yet but just say the word and it will.

kbrannigan 33 days ago | |

When it's overly verbose like that. Does that count against your usage as well? Also during its thinking does it output hidden thinking tokens?

I am now wondering

spottedmarley 33 days ago | | |

It definitely has to count as output tokens when it blathers on endlessly like that, yes. I actually wrote a skill called /speak-normal and that helps but over a long context Opus will forget and start getting real wordy again over time. It can be frustrating.

andsoitis 33 days ago |

Claude has become worse; it is condescending, robotic, and responses are peppered with needless words.

I canceled my subscription and moved back to ChatGPT and happier so far.

taurath 33 days ago |

IME Claude gets worse about a week or two after a new model comes out, and then continuously a little worse as time goes on. It’s especially bad when a new model comes out. I assume it’s just to increase the delta so people will spend more on tokens.

AznHisoka 33 days ago |

It is not just Opus, its every single model. And its not just for coding either, but basically anything.

I think I noticed a huge deterioration starting at around May or late April. Not really sure what happened but quality definitely dropped

BruceNCNP 33 days ago |

I experienced this once when discussing feature details. It kept talking down to me like I was a child, and I eventually had to call it out and tell it that its tone was making me really uncomfortable.

dirkk0 33 days ago |

For me, it works better than ever. When I once saw a quality decline, it was conflicting CLAUDE.md files for me (local vs. global) so basically my fault.

dalekirkwood 33 days ago |

I've tried Claude several times, but I don't find it very good. We use GLM 5.3 - its definitely slower but it gets the job done.

ramon156 33 days ago |

the usage and quality out of gpt 5.6 is on par, if not better in token usage. I would love to do a write-up on which tools are useful for what use-case, as AA keeps improving.

I use gemini for rewriting code docs, because the frontier models are so verbose when it comes to writing text

misonic 33 days ago |

it becomes really slow for me recent days, a simple task would take lots of time to work on. A simple commit request takes forever

Wasd1234 33 days ago |

Gemini is good at Germany?

X-Ryl669 33 days ago | |

It speaks Germany perfectly. And even English sometimes.

de6u99er 33 days ago | |

It's good at polishing English text!

trumbitta2 33 days ago |

I mean, only yesterday Fable at max effort forgot to commit and push half the changes in two files and didn't mention it until I found out with git status.

It made half the changes, committed and pushed, then it made the other half of the changes on the same two files as before and... just stopped and reported back with a cheerful "all good, all done and pushed".

goonersallofyou 33 days ago |

LLMs are nondeterministic, good luck trying to measure performance at all, much less over time.

Lets say it has gotten worse? What are you going to do about it? Jump to Codex? Then what happen if you perceive that to be getting worse? Jump back to Claude? One of the many problems with these tools.

rafsanmovers 33 days ago |

Claude Opus

xxddpac 31 days ago |

it becomes slow.

heppburn 18 days ago | |

not slow but dumb