I miss the old Claude Code

The author recounts how Anthropic earned their loyalty over time. Starting with Open Router and local models, they noticed Anthropic's models gave consistently more concise responses that got to the point quicker, drawing them toward Sonnet over comparable OpenAI models. Later, they tried summarizing a book with Gemini Pro and got a short, mediocre result, then ran the same task on Opus, which spent several minutes and produced exactly what they'd hoped for.

Anthropic's Claude Code then became a key tool. After trying Aider, Cursor, and Copilot without success, Claude Code felt like "an extension of my brain"—it would grep the repo, read relevant files, and search further as needed, exploring the codebase the same way the author would.

But the author senses a shift. While still valuing Claude Code and using it via an enterprise subscription, they increasingly have to fight the tool, which feels far less focused than before. They share tweets from others echoing the problem: Peter Gostev reports mining tens of thousands of real Claude outputs and finds Opus 5 writes very differently—longer answers, more stock phrases like "load-bearing," more em dashes, and more clauses (if, while), though not longer words. Another user, Chubby, says complaints about Opus 5 being lazy, sloppy, and verbose are consistent and match their experience, and criticizes Anthropic's response as part of the problem.