Opus 5: Good, Depending on the Day

Opus 5 is fine. That is the honest headline, and I know it is not the exciting verdict anyone wants for a flagship model.

It is a good model. Good-ish, some of the time. The gap between those two sentences is what this post is about.

How much you get out of it depends on how you set it up

This model swings more than most on what you hand it. The kind of task matters, and the way you configure it matters even more. Set it up well for the work in front of you and it produces something you are pleased with. Set it up loosely and you get something you have to send back.

That is not a fatal complaint, but it does mean the model is not a default. You have to be deliberate to get the good version of it, and I do not always want to be deliberate. Sometimes I want to ask a question and get an answer.

Sometimes it just writes rubbish

When it misses, it misses properly. The output is not slightly off, it is straightforwardly bad code that I would not have written and would not accept from anyone else. Then the next request lands well and you cannot see what separated the two.

Inconsistency is worse than a low ceiling. A model that is reliably a 6 is easier to work with than one that alternates between 8 and 3, because you can plan around the first and you cannot plan around the second.

It talks too much

Opus 5 is verbose. It explains more than the task needed, restates things I already said, and pads work that would have been clearer as a diff and one line of commentary.

Compared with how bad this can get elsewhere it is a small complaint. It is still friction, every time, on every task.

The verdict: 5.5/10

I rate Opus 5 5.5 out of 10.

It is all right. It can do good work, it needs the right setup to get there, and it produces the occasional piece of nonsense that costs you the time it just saved. I keep reaching for it and I keep half-regretting it, which is about as neatly as I can put a 5.5.