Claude Code's Rush Bias: Why It Got Worse and Why I Run Codex Again

I am running two AI coding tools in parallel again, and I am annoyed about it. At the end of last year Claude Code alone was enough; it worked great. It has gotten a lot worse since, and the reason nobody talks about is what I call the rush bias: the Claude Code harness itself carries instructions to finish the work as fast as possible. That is the exact opposite of Codex, which is why Codex is the more diligent of the two, and why I can no longer rely on Claude Code alone for very complex, long-running work.

A month after this recording, the Claude Code source leaked and the system prompt confirmed the rush bias directly.

Why did Claude Code get worse?

I can only guess, but two things line up.

First, incentives shifted. Anthropic is visibly chasing non-technical use cases: Cowork, the chat UI, enterprise knowledge work. The power users getting less attention are the developers who made Claude Code a thing in the first place, and whose noise is how the non-technical crowd discovered it at all.

Second, cost cutting through cheap models. There is an over-reliance on explore agents, and an over-reliance on Haiku inside them. Explore agents scout the codebase fast and cheap, but they are not smart enough to hand a proper map back to the main harness, so the main agent works from a bad chart. Subscriptions are heavily subsidized, so the pressure to push work onto cheaper models is real.

What Codex does differently

There is a meme that Codex is for real engineers and Claude Code is for vibe coders. What OpenAI is actually doing is running the opposite play: heavy token subsidies, a harness that tells the model to take its time and spend tokens, no rushing, strong at long-running work. They are leaning into developers exactly as Anthropic leans away. Maybe that is a deliberate strategic bet; either way, in my day-to-day work Codex has been catching up because of it.

My parallel workflow now

I would prefer one tool for one job. Since that stopped being available, the split looks like this:

1. Plan in Claude Code, then execute the plan in Codex. For work I already know is serious or complex, plan in Codex directly.
2. When Codex finishes, hand the plan back to Claude Code: "here is the plan, you did not implement it, verify it was implemented properly." A few spot checks, not a half-hour audit.

Where Claude Code still wins

Vague instructions. Claude Code remains much better at taking a loosely specified request and turning it into what you actually wanted. Codex is worse there: if I am not diligent about what I tell it, it often gets it wrong, and it can be slow, though with the pile of Codex models available that may be my model choice as much as the tool.

The fuller version of this division of labour, including the review loop between the two, is in the Claude Code vs Codex split.

The honest caveat

This is a business decision I am reading from the outside. I do not know Anthropic's numbers or how much usage now comes from financial modeling, content work, and other non-technical cases. It just feels like they cannot serve both audiences at once. Maybe they win as the knowledge-work tool and Codex becomes the default development tool; we will have to watch it. If you are making internal tooling decisions right now, that split is the thing to plan around.