|
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? |