Is Claude getting worse (at least compared to Codex)?
A trending question, a 100-hours-vs-20-hours comparison, and my own recent experience that Codex has been unusually solid.
"Is Claude getting worse?" is trending on X right now, always next to the same comparison: Codex.
I cannot fully judge "worse" — no baseline, no controlled test. But my own recent sessions line up with the mood: Codex has felt very solid lately. Fewer surprises, fewer moments where I have to stop and re-steer.
The most useful data point I found was a Reddit comparison: about 100 hours with Claude Code against 20 hours with Codex, same 80,000-line Python/TypeScript project, 2,800 tests. Their read: Claude is faster and more interactive, but needs more supervision — it can drift from CLAUDE.md and leave a migration half-finished. Codex is slower and more methodical, but treats AGENTS.md as a stable constraint and will pause to re-check its assumptions unprompted.
That is not "Claude got worse." It reads more like two different tradeoffs: a fast, chatty collaborator that needs a hand on the wheel, versus a slower one that holds its own guardrails better.
Two things I'd watch before drawing a firm conclusion:
- Model vs. harness. "Claude" conflates the model with Claude Code's agent loop, tool use, and context handling. A regression in the harness can look exactly like a regression in the model.
- Survivorship in the discourse. People post when something breaks, rarely when it quietly works. A trending complaint is a signal, not a measurement.
For now, my working move is boring: run both, on real tasks, and let the CLAUDE.md/AGENTS.md discipline point tell me something concrete — am I letting Claude drift, or is Claude drifting on me?
