I left Claude for months. Opus 5.5 is why I'm back

The Great AI Migration

For a few months, I had completely written off Claude. If you have been in the trenches building products lately, you probably know the feeling. One week an LLM is your best friend, and the next, it feels like it has undergone a lobotomy. It stops following instructions, gets overly chatty with disclaimers, and starts hallucinating code that wouldn't compile in a sandbox, let alone a production environment. I stopped paying the subscription and went back to my old reliable setups. I needed tools that worked, not tools that I had to babysit.

But the buzz around the latest updates—specifically the jump to what we are seeing in the 3.5 Opus and Sonnet ecosystem—forced me to give it another look. I am not interested in benchmarks or those cherry-picked charts the marketing teams love to tweet. I care about whether it can handle a messy SVG, manage complex frontend states, and actually act as an agent rather than just a glorified auto-complete. After a week of testing, I am back. Not because of brand loyalty, but because the utility has finally caught up to the hype.

The Frontend Reality Check

If you are building interfaces, you know the biggest bottleneck is often the bridge between design and functional code. Most AI models are decent at generating basic HTML and CSS, but they fall apart when you ask for something custom—something that actually looks like a human designed it. When I was using other models, I found myself spending more time fixing the styling than if I had just written the CSS variables myself.

Revisiting Claude showed a significant shift in how the model understands visual hierarchy and modern styling frameworks. When you ask it to build a component, it isn't just dumping a list of Tailwind classes; it seems to have a better grasp of the intentionality behind the UI. For builders, this is the difference between a prototype that looks like a 2010 template and one that feels like a modern SaaS product. It is less about the code it writes and more about the refactoring it suggests. It handles the nuance of responsive design without being prompted ten times to 'make it mobile-friendly.'

Solving the SVG Headache

SVGs are a particular pain point for developers. They are code, but they are also visual assets. Most LLMs struggle with the math required to manipulate complex vector paths without breaking the image. In my recent tests, I found that the new Opus iterations handle vector logic with a precision I haven't seen elsewhere. You can feed it a raw, messy SVG export and ask it to optimize the code, add animation paths, or change specific components, and it actually stays within the bounds of the coordinate system.

This might seem like a niche use case, but for a founder trying to move fast, being able to iterate on iconography and complex data visualizations without opening Figma every five minutes is a massive productivity gain. It is a sign that the model is finally moving past 'guessing the next word' and into 'understanding the underlying structure.'

Moving Toward Agentic Tasking

The term 'agent' is thrown around a lot in the AI space, usually to describe things that don't actually work. To me, an agentic task is something where I can give a high-level goal and the model can navigate the steps to get there without me holding its hand. Previously, Claude felt too timid for this. It would stop and ask for permission or get stuck in a loop of apologies.

The latest version feels more decisive. When I set it to task on a piece of logic—say, mapping out a data flow between three different APIs—it doesn't just give me the first step. It anticipates the edge cases. It warns me about rate limits. It suggests a retry logic that I didn't explicitly ask for. This is the 'builder-first' mentality I look for. I don't want a subservient chat bot; I want a technical partner that is looking for the holes in my plan.

Why the Skepticism Remains

Despite my return to the platform, we shouldn't get comfortable. The history of these models shows a pattern of 'model collapse' or silent updates that degrade performance over time to save on compute costs. Just because it works today doesn't mean it won't be broken by next month's update. As founders, we have to stay platform-agnostic. My return to Claude is based purely on the current output quality. The moment it starts prioritizing 'safety' disclaimers over technical accuracy, or the moment the latency makes it unusable for real-time coding, I will leave again.

The value of an AI tool is measured by the time it buys you back. If you spend that time auditing its mistakes, the tool is a liability, not an asset.

What This Means for Founders and Devs

For those of you who left Claude during the dark ages of its middle-period, it is time to reassess. The current state of the model is particularly strong for those working in the React/Next.js ecosystem or anyone doing heavy lifting with custom visualizations. It has found a sweet spot where the reasoning capabilities actually translate into usable code.

However, don't just copy-paste. The trap most builders fall into is trusting the output too much because it looks clean. The logic is better, yes, but it still lacks the context of your specific business needs and long-term tech debt. Use it to build the components and the logic blocks, but you still have to be the architect who fits them together.

Takeaway for the Week

• Audit your stack: If you haven't tested the latest Opus or Sonnet iterations on your specific codebase in the last 30 days, you are working with outdated assumptions.
• Focus on SVGs and UI: Use the model's new visual-spatial reasoning to handle the tedious parts of design-to-code translations.
• Stay lean: Don't bake your workflow so deeply into one model that you can't switch when the next competitor inevitably leaps ahead.

The AI space moves too fast for loyalty. Right now, Claude is earning its keep by being the most practical tool for people who actually build things. It isn't perfect, and it is still prone to the occasional hallucination, but the gap between 'idea' and 'execution' just got a little bit smaller.

Read the original at Lenny's Newsletter →