# Astra has regressed to Opus 5.0 levels of stupid, just without all the claudisms. — retained record

> Retained reddit record from reddit:OpenAI.

Canonical: https://highsignal.app/data/records/d015a46556c0e55c

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:OpenAI)  published 2026-10-27 20:05 ingested 2026-10-28 02:36

[original source ↗](https://www.reddit.com/r/OpenAI/comments/1xh36yw/astra_has_regressed_to_opus_50_levels_of_stupid/)

## retained text



```
Launch Astra was amazing.

In-between Astra was great.

Current Astra genuinely reminds me of bashing my head against pre-5.5 Claude models, wondering how a flagship could possibly be this dumb.

inb4 people whose "experience" is limited to the most bog-standard code generation on earth and telling it to create a "personal brand" website with zero direction coming here to tell me it's a "prompt issue".

Astra currently: Forgets what you told it, what the task was, what the instructions were. It over-generalizes, over-compresses, and mangles your instructions, notes, and requirements into something that's almost guaranteed to produce bad results when referred to later to do work. It has that annoying Claude pattern of knowing exactly what it did wrong, and what it should have done, when asked, and then never doing it.

At launch, if there was an issue with workspace permissions or Codex config, Astra would instantly be there like "here's what's happening, why, and what can be done to fix it" and then it would fix it.

Yesterday? I spent like, an hour watching it repeatedly ***do online research*** and fail to fix an issue with Git permissions locally. It eventually stumbled into a fix, in the form of a conflicting bad update a different agent did to the config at some point, and that resolved it, but I'm 100% positive that Astra at launch would have done all of this automatically without having to repeatedly do this stupid pattern of "nope, that didn't work" "you're right, here's what happened, here's what I think needs to happen next [END TURN]" "ok, so fix it?" x20.

It also suggested just running it with sandboxing/controls completely disabled instead of just resolving the permission/setting issue.

"hur hur hur but you could just go online and read the docs and configure it correctly yourself" - yep, sure could, and definitely not gonna, just like 99.9% of other people using it.
```



document fetched 2026-10-28 00:28 · canonical https://reddit.com/r/OpenAI/comments/1xh36yw/astra_has_regressed_to_opus_50_levels_of_stupid
