My thoughts on Claude Opus 4.6

Is Opus 4.6 Actually Good? A Hands-On Review

I’ve been using Claude Code daily for the past few months to build AskLexy, where I handle most of the backend and a good chunk of the Next.JS web app and Flutter mobile app, and LifeXP, where I work with same stack. I’m not reviewing this from a benchmarks perspective. I’m reviewing it as someone who relies on these models to ship real features, debug production issues, and move fast. So when Anthropic dropped Opus 4.6, I put it through the same workflows I use every day. Here’s what I found.

What’s Different in Opus 4.6 from Its Predecessors

The new Claude Opus 4.6 improves on its predecessor’s coding skills. It plans more carefully, sustains agentic tasks for longer, can operate more reliably in larger codebases, and has better code review and debugging skills to catch its own mistakes. And, in a first for our Opus-class models, Opus 4.6 features a 1M token context window in beta. Opus 4.6 can also apply its improved abilities to a range of everyday work tasks: running financial analyses, doing research, and using and creating documents, spreadsheets, and presentations. Within Cowork, where Claude can multitask autonomously, Opus 4.6 can put all these skills to work on your behalf.

They claim it’s better in coding, especially when working with larger codebases, given its larger context window. It can manage subagents better and longer, and it has got better planning, code review, and debugging skills. More details can be found on Anthropic’s blog. So let’s get to the point, “Is Opus 4.6 better than the older Opus models? Is it better than its competitors in practical situations?”

Opus 4.6 in Claude Code

I have tried various tasks with Opus 4.6, from giving it a vague prompt and asking it to make a feature, to giving a well defined product requirement document. Claude models in general are good at following instructions, and Opus 4.6 is no exception. The model (or all models in CC) in general started working faster than earlier. It did come with a cost though, it was less accurate than before. Usually if I miss something in the document I provide, Claude would clarify things if it had to. Around the time CC started acting faster, it would do so without proper understanding of the codebase. It started asking questions more frequently than it used to, questions that I had already mentioned in the document and that it could understand by checking a specific file in the codebase. It would come back asking how to implement something without (or even after checking) properly understanding the context. Once I clarify things, it goes back to a bit of thinking and reading code and then starts to write. Obviously it does write code well, but compared to Opus 4 (note: not Opus 4.5, I did mean to say Opus 4) it isn’t better. It’s just as good as Opus 4.5 in most coding tasks, maybe because of the lesser thinking and reasoning. Once it is given more time and better context, in theory, it can outperform other models.

When it comes to spawning and managing subagents, it does a good job (though it doesn’t seem to have improved) compared to its previous versions. It is still slightly better in this aspect than Sonnet 4.5. In a practical scenario, when given a detailed task specification, the difference would be negligible given the token usage, though Opus 4.6 is more “mindful” in managing the subagents.

When given a vague prompt, it does give a semi-done output, and we would have to perform more iterations to correct it (or do it manually, especially if it’s frontend and you have to correct the spacing or padding of your divs). Gemini 3 Pro still seems to perform better when given a frontend task given its better spatial awareness and alignment. Given the above drawback about CC taking less time to reason and do a specific task, and with the release of Codex 5.3, it tends to outperform Opus in a complex or huge task.

Enough about coding, what about day-to-day tasks? Is it any good at doing research or answering from a document?

Opus 4.6 in the Chat Interface (claude.ai)

Let’s be real here, Anthropic did optimize their model to perform better in their interface. For a lot of regular uses like answering daily questions, finding things from the web, and so on, it’s still good. But is it needed for these? Not really, Sonnet can very well do that. When it comes to math and asking about coding, it is better than Sonnet. I tried this by integrating a large codebase from GitHub and asking it to generate a redesigned version of a homepage with a different focus, mentioning what it should focus on and some files it could go through to get started. Its response was fast and good for the first iteration. The interface did look a bit cluttered and some things could’ve been aligned better, but in another iteration specifying changes, it was all fixed. For a general agent with a “not-so-properly” written prompt, it did good.

It has good cross-chat memory, better than ChatGPT or Gemini. It recalls things whenever needed (though it does that a bit more than needed at times). If you are a developer asking Claude for suggestions, it would mention your older projects at the end of the message asking if you want suggestions for those as well. In a way it’s good, but it could get annoying. It doesn’t overdo it though, it knows when to ask that and what the user’s intentions are. If the user wants to continue, it continues asking follow ups about the same project. Given its massive context, it remembers even a long chat about the most random thing you mention pretty well. Switching between contexts is pretty good within a chat, but I would suggest not overdoing this if you’re trying to achieve something important as it would definitely forget older messages and context sooner or later. Usage of connectors is pretty good as well, though it hasn’t seemed to have improved from its predecessors.

Is the 1M Context Window a Gimmick?

Opus 4.6 is the first in-class model to have such a massive context window. Before answering if it’s real or not, let’s understand what a context window is and why it matters.

Think of a context window as the model’s short-term memory during a conversation. Every message you send, every response it gives, every file or document you paste in, all of that occupies space in this window, measured in tokens (roughly speaking, a token is about three-fourths of a word in English). Once you exceed that limit, the model starts “forgetting” the earliest parts of the conversation, meaning it can no longer reference or reason about things you discussed at the beginning. This is why context window size matters so much, especially for developers working with large codebases or anyone having long, multi-topic conversations. A 200K context window (what most models offer today) sounds like a lot until you paste in a moderately sized repository and realize the model has already used up most of its memory before you even ask your first question.

So to answer the question in the heading, is it a gimmick? It isn’t objectively possible to answer this question, but when compared to other models (especially ChatGPT on web and Gemini on web) Opus 4.6 does have much better memory and recall. Given that it’s still in beta, it’s pretty good! It remembers the order of mentioned or specified tasks, remembers requirements, even minute details about the codebase like colors used in the app. When talking about stocks or other things, it would recall a stock mentioned in chat way above that even you would’ve forgotten (if you have a goldfish brain like I do. If not, congrats! Your context window is more than what Opus 4.6 has!!). Practically it does it pretty well.

Reasoning Depth vs. Speed Tradeoff

Now back to Anthropic’s product I use the most, Claude Code 😅

Is the model being faster any good? In reality, what matters the most is the output being precise and less hectic to work with. Especially for developers who sit and read a lot of code and documentation, having fewer worries about code not working properly is something really good to have. Earlier, even though Claude was slower, its performance was outstanding and unbeatable. Now it has come to a point where OpenAI’s Codex 5.3 and Opus 4.6 are on par, and if anything, 5.3 performs better in some cases. Given the token consumption, Codex 5.3 might be a better model when it comes to difficult tasks where depth and accuracy matter the most.

Earlier I could blindly trust Claude with its code, but recently even upon shallow testing there are quite a lot of bugs and reworks that it’s leaving behind. If I give it a big feature, it would leave 10 bugs or fixes that I have to go back and fix myself or ask it to fix. This might not (and mostly is not) be a model thing, rather it’s a thing about the tool itself, as the web interface is really good with Opus 4.6. Given the dynamic releases in Claude Code and the model itself being newer, it would take time for the team to make the tool stable with the new model. Still, for a company that’s built its reputation on reliability, shipping it in this state is disappointing.

But let’s be real, a model would do the work if you specify properly. Whether it’s Codex 5.3 or Opus 4.6, if it’s given proper requirements it would follow them. Codex 5.3, given its better reasoning, might still finish it slightly better, but it would be rare that even that happens, given the requirements are clear and less ambiguous.

So Is Opus 4.6 Any Good?

Yes! It is a good model. Given its really huge context window, it has got good potential. The only thing is that it should work better on Claude Code, which I hope would happen sooner. Till then it would be on par with Codex 5.3. Ideally, if the tasks and requirements are clearly specified to the model, it still gets the job done. And since it has a long memory, it works really well for building huge features that are well simplified (given the current state of the CC tool and the model). If code correctness is what matters to you, Opus 4.6 isn’t your thing unless you give it a well written product requirement document. And if you do write a good PRD, you could as well use Sonnet 4.5 for most of the coding tasks since it performs well for the amount of tokens it uses. I strongly recommend the latter if you are a developer who knows what you are doing. So we have to wait till Anthropic improves this database of a model to know its full potential in coding.