Claude Opus 4.5 changes everything

Experimenting with AI-Generated Code in 2025

Before I begin, I would like to clarify my position. I'm one of those people who believe that AGI will happen. I don't know when, but I believe that while the human mind and consciousness are extraordinarily complex, they are ultimately governed by the physical laws of the universe and can therefore be simulated. Someday AGI will be a reality. How far are we? I have no idea, and I don't care.

What I do care about is how the current growing capabilities of LLMs will impact what has been the source of income for my family for the last 15 years. So I have been keeping an eye on emerging tools, workflows, and new approaches to software engineering. For over two years, I have been using GitHub Copilot extensively with multiple models, coding agents, and custom agents, and for the most part it has been hit and miss.

I usually ask GitHub Copilot to implement a feature or fix a bug using the chat or coding agent, and a lot of times it would go very wrong. I have developed the habit of staging changes before each prompt, code reviewing changes for each prompt as I go along, and rolling back via git when I'm not happy with the solution. Working like this for a while means that I have been able to develop a sense of what kinds of things will work and how to break problems into steps that make it more likely that the AI agent will do what I expect.

I know some developers feel like AI agents are taking the joy of coding away from them, but I do not feel that way because it has allowed me to spend more time developing features. I feel like finding the root cause of bugs (even when I end up fixing them manually) is one of the things that LLMs can be very good at. As a result, I find myself spending way less time debugging.

Another thing, in my opinion, that is better is that I have to deal less with "repetitive tasks". After 15 years in the sector, I find that these days I enjoy spending more time trying to understand the business problem and designing a solution than actually implementing it. Once I have designed the API contracts, boundary contexts, database schema, etc., the implementation becomes very much grunt work—something that I feel I have done so many times that it is no longer enjoyable. So in 2025, I spent much more time doing code reviews and technical specs (for the LLMs) and less on implementation.

Do I feel like I have become more productive in 2025? Not really—maybe a little bit, but not a lot. I would say overall, a task took more or less the same time, but I spent more time thinking about the problem and doing verification than doing implementation.

Experiencing Claude Opus 4.5 for the First Time

It was at this point that the Christmas holidays arrived, and I happened to have unlimited Claude Opus 4.5 tokens for a couple of weeks. So I decided to work on some old side projects that I never had time to finish.

One of the projects was quite obsolete; I had not worked on it for a very long time. It originally used Create React App, but I knew that it had since been deprecated, so the first thing I did was to ask Claude to plan a migration from Create React App to Vite. I was quite impressed by the quality of the plan, so I asked it to go ahead, and 3 minutes later, I had everything working as expected.

I also had plans to make the web app work as a mobile app via Capacitor and as a desktop app via Electron. The problem is that on each platform I would have to use different native APIs—for example, to store user progress (the app is a game). On the web I use local storage, in Node.js the fs module, and in mobile apps SQLite. I needed Opus to implement an interface and implementations for each platform, then implement the builds for web, Windows, Mac, Linux, Android, and iOS. In about 10-15 minutes I had a working version. I tested it and encountered some small issues. I wanted the app to be full screen, and in Electron it was not using the entire screen, but after one or two prompts everything was done.

I was already quite impressed because when I reviewed the changes I didn't spot anything too bad. These were complex tasks, and Claude Opus 4.5 was getting them done one prompt at a time. The most impressive part is that I didn't have to explain how I wanted it done. I explained why I needed something and let the planning agent do all the planning for me. Now I only had to spend time doing code reviews, and because the code was fairly good, I was moving very fast.

I started to ask for features, and I was able to get over 10 features in one morning. I also tried Claude on a project really outside of my comfort zone. I wanted to migrate an OpenGL app to WebGL from Qt to TypeScript. I was able to implement the migration just the way I wanted in half a morning. Suddenly, when you are running 6 AI agents in parallel, it is like freaking horizontally scaling yourself. While the agents implement a set of features, you review the previous set and spend some time thinking about the next set. You feel like you are truly being much more productive.

The holidays passed, and on the 31st of December at midnight my unlimited tokens came to an end. It is now January, and when I tried to code without it, I felt like the other models seemed dumb in comparison. The joy of developing a product for me is not in coding the product—don't get me wrong, I love coding—but the true joy comes from seeing people enjoying it. With Claude Opus 4.5, I can deliver more products and more features than ever before. I can focus on listening to my users and not have to suffer all the repetitive and tedious stuff. Will the code be as good as NASA's or as beautiful as poetry? No, but it will certainly be good enough to delight users, and for me, that is enough.

I'm now looking at the higher-tier Claude subscription and thinking to myself that, considering what I have experienced, it is actually reasonable. I think Claude Opus 4.5 changes everything, and it makes me both very excited and very anxious. Excited because now developers will be able to build things that were not possible before because they required too much effort. This is going to be particularly noticeable in open source. We will soon see really powerful open source solutions (that are not just libraries) that can compete with big SaaS players. Anxious because it is hard to see how this is not going to impact job security in the long run.

Note: What I’m describing above is not vibe coding. I still review every change, work in feature branches, run CI/CD pipelines, and understand the code that ships. This is agent-driven software engineering, not blind prompt-and-pray development.

Have you tried Opus? What is the most impressive use case you have experienced?

In my next post I will talk about how I plan to combat my anxious thoughts about my career as a software engineer and focus my energy on the exciting ones.