Codex or Claude Code: the choice explained for someone who will never open a terminal

Two coding agents with the same entry price, two ways of counting consumption, and a French question with no answer online. What each does well, what neither requires, and why the choice should remain a setting.

Codex or Claude Code: both use a $20 monthly subscription for their mainstream offering (OpenAI and Anthropic pricing pages consulted September 2, 2026, with Claude Code falling to $17 on annual payment). Both launch in a terminal. The query 'codex ou claude code' currently returns no French page, although anyone having software built faces the question.

What Codex does well

Codex executes a task end to end in an existing project, runs tests, fixes failures, and repeats. Its strength is its two modes: the command-line version works on your machine's files, while the cloud version runs the task on OpenAI servers and returns a proposal to review. It has no separate price, living within ChatGPT plans: free for a trial, $8 for Go, $20 for Plus, from $100 for Pro, and $20 per user for teams billed annually. If you already pay for ChatGPT, your coding agent is included, an argument Maestro cannot beat.

What Claude Code does well

Claude Code works on your machine and files, asking permission before sensitive actions. At the time of our review, its data-use documentation was the best written on the market: it states what goes over the network, where session transcripts are stored, how long they remain, and how to disable technical reporting. On a personal account, retention can reach five years depending on a setting few people have opened. We detailed this point and the terminal's six hurdles in Claude Code or Cursor for a non-developer.

Two ways to count consumption

The often decisive detail no homepage highlights: at OpenAI, Codex and ChatGPT conversations share the same rolling five-hour window. Having software built therefore uses your chat quota, and a long conversation eats into what remains for building. An internal unit expressed in credits is consumed at 5–30 per message depending on the model. At Anthropic, the pricing page gives no numerical limits and describes multiples of Pro. In both cases, you learn what you consumed afterward, not beforehand. We explained this mechanism in token prices fall, bills rise.

What neither requires

Both write a specification if asked, break work into steps if asked, and write tests before code if instructed every session. Requiring it, retaining the document, and approving before continuation remain your responsibility, project after project. Codex offers per-action approval modes and change review; Claude Code offers per-tool permission: technical safeguards without a gate before a document you accept or reject. Codex's latest published release notes added search shortcuts for Vim users, which tells you whom the tool addresses.

Where your data goes

At Anthropic, everything is written: training controlled by a setting on personal accounts, exclusion under professional terms, and thirty-day standard retention. At OpenAI, exchanges go to their servers in both modes, even when the command runs on your machine; retention policy and storage location were not publicly documented in our September 2, 2026 review, and we will not invent them. For a practice, association, or business handling customer data, this documentation asymmetry weighs more than a three-dollar difference. The full topic is on your data.

The choice should be a setting

Maestro makes four assistants speak the same format, Claude Code, Codex, Cursor, Mistral Vibe, and allows switching mid-conversation, including midway through a project. The same scoping, breakdown, and verification work moves between engines according to pricing, outages, or confidentiality requirements. The screen looks like anything but a terminal, and documents and code stay on your Mac.

Where to start

Answer one question first: have you ever opened a terminal without stress? If yes, use whichever subscription you already pay for and add nothing; you will move faster without an extra layer. If no, the comparison changes to what surrounds the agent rather than the agent itself. The pitfalls to know before starting are in the pitfalls of AI development.