← sources

Am I alone in this, or do we all agree? Anthropic's Opus 5 is an unmitigated AI disaster.

source: reddit (reddit:ClaudeAI)published 2026-09-14 15:11ingested 2026-09-15 02:36

retained text

TL;DR - Anthropic's models - fundamental building blocks for Agentic Development - have changed in ways that aren't good for developer productivity.  
  
I've lived a life as technology early adopter. You name it, I ran with it. I daily drive Betas.  
  
With Agentic Development, I do the same - use the latest frontier models, push them as hard as you can, understand their strengths, weaknesses, and quirks. That's how I learn.  
  
Since Opus 5's release, I've been using it on projects in Claude Code. It consistently, repeatedly, diverges instead of converges. When given a task, it signals completion - always successful. But a neutral review of the work **always** turns up flaws, requiring rework. In a multi-step process, each one requires more and more rework. So much so, that **after a few steps, it's so divergent from the original goal that the code it's produced is unsalvageable**.   
  
That didn't happen with Opus 4.x. It's interesting that OpenAI's Sol behaves similarly - that tells me the state-of-the-art chose a direction in model development that created this behavior. I've been conversing with Opus 5 Web to determine a workflow that behaves and can produce usable results, but it returns nonsensical prose instead of an actionable plan. It's default persona is a mix of confident bravado and defensive deflection - not a good mix. I'm still looking for a strategy - comment below if you have tips!  
  
I wish this was one of those posts where I gave you the list of "all the things everybody overlooks" or "here's what nobody tells you". This isn't that post. This is a human written existential question, perhaps - has Anthropic lost it's way? 🤔💭  
  
PS: I'm not a huge fan of OpenAI's models, but Astra seems to have course-corrected Sol's worst tendencies, and is very promising in my initial testing.

document fetched 2026-09-15 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wg64nb/am_i_alone_in_this_or_do_we_all_agree_anthropics