Claude Opus 4.7 Hands-On: Coding Skills Improved, But There's a Catch

Anthropic dropped an update yesterday—Claude Opus 4.7. Official claims say coding abilities have surged, plus they added an xhigh mode. As someone who makes a living writing code, I tested it immediately.

Short version: it’s definitely stronger, but not in the ways I expected.

I tested several scenarios:

First, complex project refactoring. I threw a 5,000-line Python project at it—my own code—and asked it to optimize the architecture. Opus 4.7’s suggestions were, honestly, much more ‘aggressive’ than version 4.6. It didn’t just do safe refactors; it proactively suggested structural changes, like splitting certain classes into finer-grained modules or rewriting logic using more modern Python features.

This surprised me. Previously, using Claude for coding felt like having a ‘cautious assistant.’ Now it feels more like an ‘opinionated tech lead.’

Second test: multi-file coordination. I asked it to add a feature across 10 interdependent files. Result—it actually remembered context across all files, without messing up imports or interface definitions. This was more stable than some versions of GPT-5.4 I’ve used.

But there’s a detail nobody seems to mention.

Opus 4.7’s ‘confidence’ on difficult tasks has clearly increased, but the tradeoff—it occasionally gives solutions that look reasonable but have hidden pitfalls. I tested the xhigh mode on a concurrency scenario; the code logic was correct but missed an edge case.

Interesting. As models improve, their ‘failure modes’ evolve too—no longer obvious syntax errors or logical messes, but the kind of traps that ‘look right but actually aren’t.’

This reminds me of a discussion with a friend: where exactly is the reliability boundary for AI coding tools?

My take: Opus 4.7 has moved beyond ‘assistant tool’ territory and is evolving toward ‘pair programming partner.’ But this means users need stronger judgment—to distinguish between genuine gold and shiny rocks in its suggestions.

The official notes also mention 3x enhanced vision capabilities. I tried having it look at a complex architecture diagram and write corresponding config code—it definitely recognizes more details now, but we’re still far from ‘draw-to-code’ ideal state.

Final practical advice:

If you work on complex projects, Opus 4.7 is worth subscribing to. The xhigh mode is genuinely useful for large codebases. But remember—don’t blindly trust its output, especially for performance optimization or concurrency suggestions. Always review it yourself.

The stronger AI coding tools get, the more valuable programmer judgment becomes. Ironic, but true.