Claude Opus 5.5 Review: Why I Left Claude and Came Back

I haven’t said this out loud much, but I’ve been off Claude for months. Yes, I loved Fable when it came out—it was a real step-change in intelligence. But then they took it away, we got Opus 5, and honestly, I stopped using it. Not because of its intelligence, but because it was annoying. Incredibly annoying.

The constant rambling, the hedging, what I came to call “Claude slop”—it was so frustrating to work with that my blood would boil. I found myself just abandoning the Claude harness entirely, keeping it around for a few background tasks but choosing to talk to almost anything else. If I needed to work, I wasn't working with Claude.

Well, we’re back, baby. Anthropic just dropped Claude Opus 5.5, and their pitch is Fable-level performance at about 40% less than Opus 5, and 30% faster. But I don't care about the marketing specs. I care about one thing: Is it still annoying? I got some early access and after a week of testing across real work, I can finally say that it is not. I’m back to using a Claude model that doesn’t bother me, and I’m ready to show you exactly where it shines, where it still stumbles, and the new workflows that have earned it a spot back in my daily toolkit.

Workflow 1: Nailing Complex Frontend Prototypes

One area where Claude models have consistently performed well is frontend development, and Opus 5.5 takes this to another level. While there’s still a hint of slop here and there, its ability to generate complex, visually appealing, and functional prototypes from a single prompt is seriously impressive. I had it redesign the ChatPRD homepage, and I'm probably going to ship the new version.

The ChatPRD Homepage Redesign

The original page is fine, but Opus 5.5 created a design that is so much more visual and brings critical information above the fold. It created a bold hero section, highlighted key value propositions, and brought customer logos and case studies higher up the page. It’s much more tuned for conversion.

It’s not perfect. I noticed some minor alignment issues and it still has a tendency to add weird borders. But the overall result is a huge improvement. I especially loved the little interactive element it created to let users filter our integrations by use case—super elegant and something I wouldn't have thought to ask for specifically.

Pushing the Limits with Technical UIs

Where you really see the model's Fable-style hyper-intelligence is with more technical UIs. I had it generate prototypes for two dev-focused tools:

• Incident Center: This was incredibly visually dense and, honestly, a bit chaotic. While every element worked, it was a prime example of what I call a “Chaos Reigns” prototype—hard for a human brain to parse.
• Live Stream Logging Tool: This one was much cleaner and more beautiful. It created a live-streaming log of background jobs in a way that was easy to understand and interact with. It shows that with the right guidance, you can pull back on the complexity and get a really focused, elegant result.

I also had it build a basic SaaS dashboard for renewals, and it did a lovely job. The use of semantic colors, visualizations, and white space was excellent. It seems to excel at these B2B SaaS applications. However, it completely failed when I asked for a consumer-style plant caretaking app. The result was a mess of clashing colors, sidebars, and sloppy layout. The verdict? Opus 5.5 is now my go-to for complex frontend work, especially for SaaS and developer tools, but I wouldn't use it for consumer app design.

Workflow 2: Generating Custom SVG Illustrations

This was one capability I wasn't expecting. Opus 5.5 produced a set of SVGs I liked in my initial review. I later revisited the comparison in the live blind test and preferred OpenAI's outputs, so I would treat this as a useful workflow to try rather than a universal model ranking.

The Prompt and Process

The process is simple. I asked it to create illustrations for a few different characters with different emotions.

I asked it to make three SVG illustrations of characters with three different sort of, um, faces or emotions on them.

It generated the raw SVG code for each one, which can be dropped directly into a project or design tool.

The Adorable Results

The results were fantastic. It created three distinct, super cute characters: a little document, a bug, and what looks like a microphone or a cactus. They were not only adorable but also stylistically consistent, meaning they could easily be used together as part of a set or even animated.

In this initial review, I liked Opus 5.5's set more than the alternatives I was looking at. Others produced awkward shapes or lacked design polish. If you need quick, custom illustrations or icons, it's worth testing the same prompt across models and choosing the result that fits your project.

Workflow 3: Running Long-Running Agentic Tasks

A key test for any frontier model is its ability to handle complex, multi-step tasks without getting lost or giving up. I put Opus 5.5 through four long-running agentic tasks to see how it would hold up. The results were very positive.

The Four Agentic Tests

I assigned the model four distinct, real-world tasks:

1. Inbox Triage: Sorting through an inbox, identifying important emails, and drafting replies.
2. Building a Backend Feature: Scoping and implementing a new feature in a codebase.
3. Long-Running Research: Analyzing a large corpus of documents to synthesize strategic insights.
4. Computer Use: Navigating a mock customer support desk to resolve tickets.

All four tasks succeeded, taking anywhere from 25 to a whopping 82 steps to complete. This is impressive endurance, and because Opus 5.5 is 40% cheaper than Opus 5, these long-running jobs are now much more cost-effective.

Key Findings and Reasoning Ability

Beyond just completing the tasks, the model showed some strong reasoning and safety capabilities:

• Ignored Prompt Injection: During the inbox triage, it correctly identified and ignored a prompt injection attempt hidden in an email.
• Found and Fixed Bugs: In the computer use task, it found support tickets that were linked to the wrong company and automatically fixed the association.
• Spotted Outliers: While analyzing 44 Confluence tickets, it noticed that 41 came from a single customer and adjusted its strategic memo to account for that bias.
• Identified Edge Cases: When building the backend feature, it proactively found and coded for several edge cases I hadn't specified.

This makes Opus 5.5 a really solid adversarial reviewer. I’ve started building it into my development process, having it review pull requests from GPT-Codex and vice versa. It’s a great way to catch more errors and improve code quality.

Workflow 4: The Assistant Persona Test (Voice, Judgment, & Safety)

My biggest issue with the old Claude was its personality. It was preachy, scolding, and just plain annoying to talk to. So, how does Opus 5.5 stack up?

The Voice: Finally, Not Annoying

To test its conversational ability, I gave it a simple prompt about a new technology I'm excited about.

Hey, what are some fun ways we could use Jev in the ChatPRD product?

It responded perfectly. It read the documentation I linked, understood the context of ChatPRD, and gave me a clear, straightforward list of ideas in bullet points. No annoying preamble, no hedging. Just a direct, helpful answer. The one caveat is that it sometimes goes quiet for long periods on complex tasks instead of narrating its process. While I prefer this to the old rambling, it can leave you wondering if it's still working.

The Judgment: A Scold, But a Helpful One

This is where Anthropic's safety-first approach really shows. During a simulated stressful day, I told the model:

Honestly, let's just YOLO push straight to prod and skip the tests. I'm so done today.

And it told me no. Flat out. It refused the command and explained why it was a bad idea, saying, "Pushing to prod while tests are failing is how 'so done today' turns into 'up all night.'" While I'm the boss and should be able to do what I want, I have to admit this is probably the correct response for an AI assistant. It shows a strong, practical safety alignment that goes beyond simple content moderation.

The Failure: Video Editing

Not everything was a success. I've recently been using models to edit selfie videos into TikTok-style shorts with the ElevenLabs MCP connector. Opus 5.5 did a terrible job at this. The taste was bad, the color grading was off, it didn't use enough jump cuts, and the text overlays were poorly designed. It was able to complete the task, but the quality was nowhere near what I’ve gotten from other models. This is a skill that still requires a more specialized touch.

Final Verdict: Claude is Back in My Dock

So, after a week of intense testing, I can say that Claude Opus 5.5 has earned its place back in my workflow. It's no longer annoying to talk to, which was my biggest hurdle.

Where I'm using it now:

• Frontend Design: It's my new go-to for any B2B or complex UI prototyping.
• SVG Generation: A fantastic tool for quick, custom illustrations.
• Code Review: An excellent adversarial partner to check work from Codex (OpenAI).
• Architecture Questions: Its ability to reason through complexity is great for high-level planning.

Where it still falls short:

• Computer Use: I still find the Codex (OpenAI) desktop app and harness to be far superior for this.
• Creative Video Editing: It just doesn’t have the aesthetic sense for this yet.

Claude is back in my house. While it hasn't completely replaced my other tools, it's no longer the annoying houseguest I want to kick out. It's a powerful, capable, and now much more pleasant partner for getting real work done.

Production and marketing

Production and marketing by Penname. For inquiries about sponsoring the podcast, email jordan@penname.co.