A few days ago, Anthropic released its new flagship model — Opus 4.6.
In my opinion, Claude models have long been the best for coding, so I was very curious to see how much better it would get.
Until now, I mostly used Opus 4.5 and occasionally relied on Sonnet 4.5 because Opus didn't have a 1M-token context window.
And that was sometimes limiting — especially with larger files or more complex refactoring tasks.
For vibe coding, I've found that keeping classes as short as possible works best in the long run.
In our codebase, we use what we call Actions for this.
Each action performs one specific task — nothing more.
We used to use service containers, but as more and more things were added to them, the files and classes became huge and difficult to work with.
I also tried Gemini 3 Pro, which I found was sometimes better at frontend work than Opus 4.5.
Opus, on the other hand, excelled more at backend work.
But with the arrival of Opus 4.6, that has changed significantly.
I didn't have much time to test it over the weekend, but today (Monday) I finally got around to it — and my jaw dropped on the very first prompt.
I decided to give it a task that required a larger context — at least 200–300k tokens.
Opus 4.5 couldn't handle it last week, and Sonnet is simply weaker, so it failed as well.
I gave it a fairly vague prompt for a rather complex refactoring task.
The project also has a lot of dependencies throughout the codebase, so I was curious to see how it would handle them.
And it handled them brilliantly.
Not only did it get everything right on the first try, but it also fixed mistakes in my existing code and pointed them out to me.
I've been sitting here “coding” with it all day.
I've burned through about $150 today.
On the other hand, I've written a huge amount of code and fixed many things I'd been putting off for a long time — the classic skeletons in the closet.
And Opus hasn't made a single mistake.
To me, this is exactly the kind of AI that genuinely makes life easier — and the kind I've been waiting for.
Honestly, I can't imagine where it will be in a few years.
You still need to “hold the model's hand” and proceed step by step, but it already feels almost perfectly polished.
I also tested GPT Codex 5.3, and in my opinion, OpenAI is currently quite far behind Anthropic (at least when it comes to coding).
If you're a developer, I definitely recommend trying Opus.
For me, it's a game changer so far.
Until now, I had to painstakingly check, explain, delete, and try everything again.
Now, for the first time, I feel like things work the way they're supposed to — on the very first try.
