Opus 5.5 is aiming at the project you keep putting off

The release, without the victory lap

Claude Opus 5.5 arrived on September 22. Anthropic describes improvements in complex coding, knowledge work, communication, and efficiency. It reports lower typical running costs than Opus 5 and output generated more than 30 percent faster. The announcement includes internal evaluations and early customer examples involving large codebases. It also cautions that benchmark margins have become a less dependable guide to differences in actual work. We have not independently reproduced the company’s performance claims.

Source: Anthropic: Introducing Claude Opus 5.5 ↗

Our take: define done first

Our practical read is that stronger models make a clear project brief more valuable. If an assistant can work across many files, an ambiguous instruction has more room to spread. For a personal portfolio, define the outcome before asking for edits: every navigation link works, images have meaningful descriptions, and the contact path reaches the right destination. Ask the assistant to explain its plan and identify assumptions before it changes anything. The brief gives you something concrete to evaluate when the model returns a confident completion message.

Make your experiment reversible

Try a copy of a small project you already understand. Choose a single improvement, such as locating broken links or simplifying repeated styles, and request a summary of every changed file. Open the result and use it like a visitor would. A polished screenshot is only one part of success; test the buttons, keyboard navigation, and narrow-screen layout too. Record the edits you still need to make yourself. That evidence tells you more about whether a bigger model earns its price than a dramatic demo on an unfamiliar codebase. Keep the original so you can compare both versions.