> When I upgraded to Sonnet 4.5, it became less often that it gave phantom answers. But it still sometimes wasn’t able to handle some complex problems. We’d go back and forth and I’d try to give it hints. But with Opus 4.5, that happens much less often now.

The real annoying thing about Opus 4.5 is that it's impossible to tell most people "Opus 4.5 is an order of magnitude better than coding LLMs released just months before it" without sounding like a AI hype booster clickbaiting, but it's the counterintuitive truth. To my continual personal frustration.

The thing is, it still feels like a mixed bag for me.

It's good enough for things I can define well and write okay code for.

But it is far from perfect.

It does too much, like any LLM. For example, I had some test cases for deleted methods, and I was being lazy and didn't want to read a huge test file, so I asked it to fix it.

It did. Tests were green because it mocked non-existing methods, while it should have just deleted the test cases as they were no longer needed.

Luckily, I read the code it produced.

The same thing happened with a bit of decorators I asked it to write in Python. It produced working code, tests were fine, but I reworked the code manually to 1/10 of the size proposed by Opus.

It seems magical, even thinking, but like all LLMs, it is not. It is just a trap.

> The real annoying thing about Opus 4.5 is that it's impossible to tell most people "Opus 4.5 is an order of magnitude better than coding LLMs released just months before it" without sounding like a AI hype booster clickbaiting, but it's the counterintuitive truth. To my continual personal frustration.

The problem is that these increases in model performance are like the boy who cried wolf. There's only so many times you can say "this model is so much better, and does X/Y/Z more/less" and have it _still_ not be good enough for general use.

What people do not understand is that this really depends on what language you target. So if I write Rust then you sound like an AI hype booster but if I write TS or Python maybe not so much.

From my experience Opus is only good at writing Rust. But it's great at something like TS because the amount of code it has been trained on is probably orders of magnitude bigger for the latter language.

I still use Codex high/xhigh for planning and once the plan is sound I give it to Opus (also planning). That plan I feed back to Codex for sign-off. It takes an average additional 1-2 rounds of this before Opus makes a plan that Codex says _really_ ticks all the boxes of the plan it made itself and which we gave to Opus to start with ...

That tells you something.

Also when Opus is "done" and claims so I let Codex check. Usually it has skipped the last 20% (stubs/todos/logic bugs) so Codex makes a fixup plan that then again goes to through the Codex<->Opus loop of back and forth 2-3 rounds before Codex gives the thumbs up. Only after that has Opus managed to do what the inital plan said that Codex made in the first place.

When I have Opus write TS code (or Python) I do not have to jump through those hoops. Sometimes one round of back and forth is needed but never three, as with Rust.

First impressions matter. I felt the same reading comments suggesting that people who praised GPT-5.2-Codex recently were shilling for OpenAI when it has actually gotten much better, faster and the most important one, more time before you reach your weekly limit.

  To my continual personal frustration.