Claude used to lay code and Codex would fix it, in a pretty fair 1:1 time split. Today it’s taking Codex four times longer to fix Claude’s code. 1:4. That’s hardly ideal.
Fixing now takes four times as long as laying, and that number decides what happens next.
The split worked because the two had different temperaments. One lays fast and leaves rough edges, the other files them down, and when the filing takes about as long as the laying the pairing pays for itself. At four to one, the fixer is doing most of the work and the layer is mostly generating things to fix.
The ratio matches a wider mood. Last week VentureBeat described users reporting that Claude “feels less capable, less reliable and more wasteful with tokens than it did just weeks ago,” with Anthropic disputing the cause.
Using 4.7. Anthropic shipped it on April 16 and says it “devises ways to verify its own outputs before reporting back” and “pays precise attention to instructions.” Pricing is unchanged from 4.6, at $5 per million input tokens and $25 per million output. So the price of laying the code is the same, and the time spent cleaning it up has quadrupled.
Haven’t tried to roll back yet, because I’d heard it failed midstream too often. Not enough compute available for 4.6.
The other half of the pair is OpenAI’s. Codex CLI is, in OpenAI’s words, “a coding agent from OpenAI that runs locally on your computer”, and OpenAI recommends signing in with a ChatGPT plan to use it. The fixing has been happening there. If Codex is the better fixer today, the obvious question is whether it’s the better layer too, with Claude checking behind it.
By the way — can confirm you can now use Claude, Codex and Gemini, all three primary labs, in OpenClaw on subscription. Crazy.
So I think I may switch: Codex laying code and Claude fixing it. Reverse the flow.
