TL;DR: After 165 sessions with Claude Code and 27 with Codex CLI, one thing is clear: use Claude for interactive, exploratory work and Codex for autonomous tasks you can delegate and forget about. It’s not about which is better — it’s about when to use each. My data shows the combination outperforms either one alone.

It’s 9:00 a.m., coffee in hand, and I have a vague idea about restructuring a module. I don’t know exactly what I want yet — I just know the current setup isn’t right.

I open Claude Code.

Not because it’s “the best,” but because I need to think out loud with something that understands context. I start with, “This service has too many responsibilities. Help me split it up.” And so begins a dialogue. It suggests a separation, I discuss it, ask for another approach, and it shows me the code changes before implementing anything. It’s a conversation.

Three hours later, the module is split, the tests pass, and the design is better. But I’m now down to 40% of my Claude quota. On top of that, I’ve accumulated a list of mechanical tasks: updating integration tests, cleaning up obsolete imports, reorganizing a directory.

I open Codex.

I give it the tasks, tell it to work autonomously, and take my lunch break.

The Data Behind the Pattern

For months, I’ve been monitoring my usage with a menu bar app that tracks sessions, tokens, and hours. Here’s what the numbers reveal:

  • Claude Code: 165 sessions, 160,893 messages, 28,052 tool calls. Models: Opus 4.5 and Opus 4.6. 5.6 billion tokens retrieved from cache.
  • Codex CLI: 27 sessions. Model: GPT-5.4. Mode: danger-full-access, approval policy: never.

The first surprise? The time-of-day distribution for Claude Code usage:

09:00    1 ▏
10:00    5 ██
11:00   10 ████▌
12:00   14 ██████▎
13:00    7 ███
14:00   16 ███████▏
15:00   21 █████████▍
16:00   15 ██████▋
17:00   18 ████████
18:00   13 █████▊
19:00   14 ██████▎
20:00    9 ████
21:00   10 ████▌
22:00    8 ███▌

70% of my Claude Code sessions happen in the afternoon and only 22% in the morning.

At first, I thought my theory of “Claude for the mornings” didn’t hold up. But when I looked at what I was doing at each time, it all made sense.

Mornings Are for Exploration, Afternoons for Execution

Morning sessions with Claude Code are fewer but longer. These are design sessions: “How should this work?”, “Review this plan,” “Suggest alternatives.” They’re iterative back-and-forth discussions where I change my mind three times before deciding.

Afternoon sessions are shorter and more frequent. Implementing decisions made in the morning. Fixing bugs. Adjusting tests. Tasks where the direction is clear, and all that’s left is to execute.

This is where Codex shines.

Why Codex for Autonomous Tasks

Codex CLI with GPT-5.4 in `approval_policy =

This article was originally published in Spanish and translated with the help of AI.