Opus Feels a Little Strange Again
Opus Has Felt Odd for the Past Few Days
Over the past few days, Opus in Claude Code has felt strangely less sharp.
I have seen similar comments appearing around social media as well.
It misunderstands context that it used to grasp immediately.
It does not fully read the design intent.
Its reviews feel shallow.
If I give very detailed instructions, it still works, but my supervision cost has increased.
It feels as if an excellent subordinate suddenly started showing up sleep-deprived.
I Do Not Think the Model Itself Degraded Overnight
I do not think the model itself can deteriorate in just a few days.
But from the user’s side, we cannot see what is happening behind the scenes.
Is it an inference setting?
Is it routing?
Is it context handling?
Or is it simply my imagination?
Honestly, I have no proof.
My guess is that it has probably been affected by the recent turbulence around FABLE.
There was also a previous wave of discussion about whether a change in thinking-level settings had reduced performance.
Since then, I have come to think:
“Well, it would not be surprising if something were being changed behind the scenes.”
I Have Stopped Getting Angry About It
Recently, I get angry about this less often.
In the past, I would have been furious.
“Why can you not do today what you were able to do yesterday?”
But now I simply think:
“I guess today is one of those days.”
AI seems to become foolish from time to time.
It is like the weather.
Today, it rains.
Today, Opus is a little off.
Some days are just like that.
Trying an Older-Looking Opus
For now, I tried a workaround that was being shared online:
/model claude-opus-4-7So far, this one feels more like the old Opus.
At least that is my subjective impression.
The latest version is not always the best.
That has long been true in the software industry.
If the latest version has problems, you roll back to the previous one.
AI may be the same.
Preparing to Move to Codex If Necessary
If even that does not work, I will move to Codex.
In fact, I have recently been preparing for that as well.
For every Git commit, I have Claude create handover notes.
What was implemented?
What was the design intent?
What issues remain unresolved?
What should be done next?
I have it summarize all of that.
If Claude suddenly becomes unreliable, I can hand those notes to Codex and continue development at a minimum level.
It is, in effect, a backup operation.
In the past, the future people imagined was:
“AI will write code instead of humans.”
The reality turned out to be different.
AI writes code instead of humans.
AI designs instead of humans.
AI writes documentation instead of humans.
And humans write handover notes for the moment when AI suddenly becomes foolish.
I am no longer sure whether this world feels futuristic or not.