Claude Code vs Codex CLI: Same $20, One Leaves a Daemon

Both cost $20 to start, so the real difference is what each agent leaves running on your machine

Published 18 October 2026 Updated 18 October 2026

Claude Code vs Codex CLI is a decision that looks settled before you make it: the same $20 entry price, the same job, two tools you could describe with the same sentence. One of them now keeps a process alive on your machine after you close every window, and the other waits to be asked. That difference, not the benchmark scores, is what you actually live with. Here is what separates them on price, on models, on defaults and on the kind of attention each one demands from you.

Claude Code vs Codex CLI: two agents, one terminal

Claude Code is Anthropic's agent. Its documentation describes a tool that lives in your terminal, in VS Code and JetBrains extensions, in a desktop app and in the browser.

Codex CLI is OpenAI's answer: open source under an Apache-2.0 license and written in Rust, with its own editor extensions and its own desktop app.

Both read your codebase, edit files and run commands when you ask in plain English. On a feature list, you could swap the logos and nobody would notice.

The $20 tie

If the two tools do the same job, you might expect the price to decide. At the entry level it does not.

Claude Code comes bundled with Claude Pro, which is $20 a month, or $17 a month if you pay for the year. Codex comes with ChatGPT Plus, also $20 a month. Same ticket, same door.

There is one thing people miss about Codex being open source: the code is free, the intelligence is not. Codex runs on your ChatGPT subscription, so the model doing the work is a paid service. Neither tool is a way to have an agent write code for you without paying its vendor.

The ladder above the entry plan is nearly the same height on both sides. Anthropic's Max plans start at $100 a month for 5x usage. OpenAI matches that step exactly and adds a tier at $200 above it. Price will not break this tie.

A scoreboard of retiring models

So does one of them simply write better code? The headline number belongs to Claude Opus 4.5. Anthropic's own announcement puts it at 80.9% on SWE-bench Verified at its high-effort setting, the first model past 80% on that test. That is the vendor's own figure, not an independent measurement, and it sounds decisive.

Then look at the calendar. That score is from November 2025, during a week when the crown changed hands in days: OpenAI's GPT-5.1-Codex-Max had taken it the day before, after beating Google's Gemini 3 Pro. A score like that is a snapshot of a race, not of the tool you would install tonight.

And those models are already being replaced. OpenAI has set GPT-5.5 to retire from ChatGPT and Codex on October 14, 2026, steering Codex users to GPT-5.6 Sol; the API keeps the old model, but your Codex sessions will not. Anthropic, for its part, released Opus 5.5 this autumn and says it matches its previous flagship on most work at 40% lower cost (again, the vendor's claim rather than anyone independent's measurement).

The leaderboard tells you who won a week last year. It does not answer the question that matters more: what does the agent do when you are not watching it?

The process that stays behind

For Codex, since version 0.157.0, the answer is that it keeps running. A GitHub issue on OpenAI's own repository documents that this version made a background daemon the default. It copies about 374 MB of files into the Codex folder in your home directory and stays alive after every CLI window closes.

It also updates itself. The daemon checks for updates every hour, after a first delay of 5 minutes. A developer write-up on turning it off adds a sharper detail: the updater can stop the daemon with only a 60-second grace period before killing it outright. Picture a long agent task working in the background when that update lands. You did not start that process, and you did not choose its schedule. The issue, filed on September 25, 2026, asks for one thing: make it opt-in.

The fix exists today. Run codex features disable daemon_auto_start, or set daemon_auto_start to false in config.toml. That is the main catch with Codex: a capable open-source tool whose newest default you have to know about before you can undo it.

Claude Code waits to be asked

According to its documentation, Claude Code hides no process like that. It has background agents and scheduled Routines, but you create them yourself, with the /schedule command or from the web or desktop app. Nothing of the kind is switched on for you.

Now the fair part. Both tools auto-update by default when you install them with their native installer scripts, so the difference is not updates. It is whether a process you never asked for stays alive on your machine.

On setup, the two are level. Both ship native installers for macOS, Linux and Windows, plus Homebrew and WinGet packages, and neither project lists hardware requirements. You will find them on GitHub, at anthropics/claude-code and openai/codex. Both move fast: Codex reached version 0.163.0 on October 15, and Claude Code is on version 2.1.30x. On paper, Claude Code wins the default, but that caution has a price.

The cost of being careful

One YouTube creator who used both tools for a week came back with no single winner. In their view, Codex has sandboxed autonomy and burns fewer tokens, while Claude Code is more transparent and asks for more review. A second comparison video, from August 2026, landed on the same split: faster and cheaper on one side, more careful and better reasoned on the other.

Those are opinions, not measurements, and they are worth weighing as such. But they point at Claude Code's real catch. Its review-heavy style asks for more of your attention per task, and heavier use pushes you toward the Max plans. Don't use Claude Code if you want to hand over a job and walk away.

So the question is no longer which tool is smarter. It is which approach you are willing to sign for: hands-off and fast, with a process that stays behind, or slower, with nothing running that you did not start.

The verdict

What decides it is the default: what each agent does on your machine when you have not asked it to do anything. Price ties at $20, and the famous scores belong to models being retired, so the default is the one difference you meet every day.

Use Codex instead of Claude Code when you want maximum autonomy, already pay for ChatGPT, and are happy to edit a config file; switch off daemon_auto_start on day one. Take Claude Code if your machine holds client code, secrets or a strict company policy, and you would rather review more of what the agent does. If you are buying for a team, check the seat prices first: Anthropic's premium seats with more Claude Code usage are $100 a month each on an annual plan.

Here is the line to take back to your team. Don't pick the agent with the best score. Pick the one whose defaults you would accept without reading them.

Our Short on this: mattpocock/skills: "Skills For Real Engineers".

Sources