I have been a long time user of Claude Code. I never felt the need to try other coding tools, and the times I did, I always came back. Across many projects, personal and professional, Claude has been my pair programmer. I have enough experience across generations of Claude models that, just by looking at the output, I can tell when a model switch has happened, strictly in the context of coding tasks. I have noticed models lean on some aspects heavily like bullet points or emojis or markdown tables or ascii diagrams or just a walls of text (4.8). Following the Claude Code 4.7 and 4.8 releases, the performance is immensely worse. Here is how I qualify "worse":

1. Coding tasks: they don't follow instructions at all. They consistently miss close to a quarter of the ask, which cascades into bloat over time. 2. Reviewing existing code: they literally fail to read the code or use the tools properly.

Both of these have, in my opinion, eroded the time saving value of the tool. Hard opinion: real code review is harder than coding.

So I tried Codex. This is my first time, so I have only a little experience with it, but the distinction is clear. Codex is remarkably precise about the exact changes I need, reliable nearly 95% of the time. What it lacks is the flair, the bombastic ideas and presentation that Claude has. I use Claude to discuss ideas, it gives me great variety and draws ascii block diagrams Codex never could. But Codex is the most reliable coding agent.

Suggestion: Don't trust Claude when it says it finished something :) Always review it, at least post-4.6.

Is this just my experience, or do others feel the same?