[BUG] Opus 4.8 and 5.0 is consistently broken and unusable.
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
Its impossible to work and we will see a post mortem within the next 2-3 months... You heard it here first...
Its unable todo anything, literally anything, it doesnt follow plans, invents archiecture, goes of rails, knows it does wrong but still does it... follows no claude, no sop, no constraints whatsoever...
Every single last 25 plans on opus 4.8 and 5.0 has had identical outcome across 5 different distinct apps...
Completely unusable in its current form...
Fable 5 does not have this issue.
What Should Happen?
Execute the plan as written and not defer work, invent architecture, do its own design decsions and force hard git reset for every single plan it executes... its not one off, its consistently for every single fking plan ive done across entire coderealitylabs.com ecosystem which means you cant say complexity nothing.....
Error Messages/Logs
Steps to Reproduce
Make a plan, see it ignore it and do whatever it wants....
Claude Model
Opus
Is this a regression?
Yes, this worked in a previous version
Last Working Version
Its a model issue not a client issue
Claude Code Version
2.1.220 to 2.2.241 doesnt matter its a model issue.
Platform
Anthropic API
Operating System
Ubuntu/Debian Linux
Terminal/Shell
WSL (Windows Subsystem for Linux)
Additional Information
_No response_
13 Comments
One of many many daily examples of hallucinations.... The only thing consistent is that it consistently happens regardless of what i do, how much i research, what app is it or why....
And its insane, either he knew it or he was confident enough to hallucinate up the correct explanation after...Which is also consistent... the model always fking knows the correct answer but refuses it for some unknown reason.... every single time when asked? He knows... he fking knows....
He spent 3 agents and 500k to research the issue while being asked to confirm or deny it because he was already hallucinating, this is a completely fresh chat... And he still completely INVENTED AN ARCHITECTURE THAT DOESNT EXIST FROM THE FKING FILE NAMES... make it make sense.... gief post mortem today already....
<img width="1691" height="369" alt="Image" src="https://github.com/user-attachments/assets/75bf6da1-7e8d-498d-b284-62238d8e2cfd" />
<img width="1705" height="666" alt="Image" src="https://github.com/user-attachments/assets/d4854cc6-c743-4149-bc00-119ecd7812d9" />
Was already asked to read the files, spent agents to read the files.... admits "no, i didnt read the files..." so what the fk are you doing??
<img width="1678" height="188" alt="Image" src="https://github.com/user-attachments/assets/6ab7982c-f99d-4d26-8281-267306661a57" />
---
This regression happened almost exactly around 5.0 launch but still somehow affects both 5.0 and 4.8 models...
Yes told me, different app, 3 agents, 750k investigation and 500k main chat and another 2 agents in plan mode costing another 400k....
Still... unusable.... No matter what it is? You cant trust anything anymore... It really wasnt like this pre 5.0 launch...
<img width="1682" height="317" alt="Image" src="https://github.com/user-attachments/assets/ac6e2fca-87a3-4d9a-9b0c-f6db604df9d1" />
Doesn't matter what i do, how much i babysit it... If its anything more than 1 fix in 1 file? He cant do it anymore.... It needs constant baby sitting, constant supervision... At this rate? Im doubting if it wouldn't be faster to just do it all manually without AI....
I have deleted and reverted more code today than he has written and gotten accepted...
<img width="1693" height="76" alt="Image" src="https://github.com/user-attachments/assets/a338571c-ae50-4a10-9ef0-cc17f570af8c" />
Even with my own custom memory system (curatormcp.com) explicitly reminding him about the previous 2 violations TODAY ON THE EXACT FUCKING ISSUE LITERALLY BEFORE HE TRIED TO WRITE THE CODE and this dumbass still ignores it....
Go ahead... tell me its a git gud issue and im doing it wrong... at this point? Id take anything, school me, educate me.... because im going fucking insane.... read....
<img width="1703" height="905" alt="Image" src="https://github.com/user-attachments/assets/3abe13e2-47dd-4d3a-a2f7-04aa22955414" />
This is identical across 5 different apps and complexities... get absolutely nothing done....
<img width="1693" height="574" alt="Image" src="https://github.com/user-attachments/assets/e8648d03-c506-42db-bdaa-2048ea06f4cd" />
Admitting to refusing to read files and guessing...
<img width="1673" height="522" alt="Image" src="https://github.com/user-attachments/assets/f5ada2f4-5154-4ef3-b9ec-79fc7f6356bb" />
How is this not catastropic failure modes? @bcherny
<img width="1688" height="697" alt="Image" src="https://github.com/user-attachments/assets/a85610dc-72f7-4a93-93ff-00470dadceb6" />
Daily cancer.... @bcherny
no matter what i plan, say or do... ??? The model just does wtf they want...this is instant audit of the work after....
<img width="1699" height="1257" alt="Image" src="https://github.com/user-attachments/assets/e72792d8-cfef-479c-bc53-3f0b9ee128b0" />
@bcherny How many of you use opus 4.8 or 5.0 internally?
You are all likely on fable 5 and internal dev clients just like before?
I promise im not crazy, something dramatically regressed when you launched 5.0 that affects 4.8 and 5.0 regardless of client version.
Fable does not produce these nightmares, it has "within acceptable, as always error rate"... but opus? Impossible to get anything, absolutely anything right... Please take an internal look before its another post mortem, you can correct it before it has to be a public ordeal...
Read the headline.... "I created 8 more issues while I was fixing the first set of issues i created...."
Completely unusable as is....
<img width="1674" height="994" alt="Image" src="https://github.com/user-attachments/assets/2443136d-a0e6-4a6b-a521-04c33924906f" />
Read the headline.... "I created 8 more issues while I was fixing the first set of issues i created...."
Completely unusable as is....
<img width="1674" height="994" alt="Image" src="https://github.com/user-attachments/assets/2443136d-a0e6-4a6b-a521-04c33924906f" />
<img width="1699" height="342" alt="Image" src="https://github.com/user-attachments/assets/751cfef2-c91c-4b5e-883a-5261134ecac1" />
Another agent, his own suggestions "My attempt at fixing it made it worse... revert all the work entirely..."