Head to head

Codex vs Claude Code: seven rounds and an honest verdict

Updated November 17, 202610 min read

Codex vs Claude Code is the choice most people face first when they want an AI to write code for them. Codex comes from OpenAI and is part of ChatGPT. Claude Code comes from Anthropic and is part of the paid Claude plans. Both read your project, change files and run commands while you watch. Below we put them through seven rounds, score each one, and finish with a clear pick for each kind of user.

The verdict in five lines

  • Codex wins on price: you can try it free in the desktop app, and it comes with any ChatGPT plan you already pay for.
  • Claude Code wins on control: more permission modes, hooks, subagents and a richer project notes file.
  • Codex wins on Windows: its sandbox works on plain Windows, while Claude Code's needs WSL 2.
  • Neither is faster in every case. Speed depends on the task and on which model is newest.
  • Final score: 3 rounds each and one tie. Your budget and the size of your jobs should decide it.

Round 1: installing Codex and Claude Code

Both tools install from one line in a terminal (the black text window where you type commands). After that you open a project folder, type the tool's name and press Enter. The first run asks you to sign in with your account.

# Claude Code on Windows (PowerShell)
irm https://claude.ai/install.ps1 | iex

# Claude Code on Mac or Linux
curl -fsSL https://claude.ai/install.sh | bash

# Codex on Windows (PowerShell)
irm https://chatgpt.com/codex/install.ps1 | iex

# Codex on any system with Node.js
npm install -g @openai/codex

# Codex on a Mac with Homebrew
brew install --cask codex

The one-line installers above do not need Node.js (a program many coding tools ask you to install first). Only the npm route for Codex does. If you would rather skip the terminal, both have a desktop app. Codex has its own app, and on Windows it comes from the Microsoft Store. Claude Code lives in the Code tab of the main Claude app.

Sign-in is about equal. Claude Code wants a paid Claude plan or an API key (a secret code that bills you per use). Codex accepts a ChatGPT account or an OpenAI API key. Full walkthroughs live in our Codex CLI guide and how to use Claude Code.

Round goes to: a tie. Both are quick, and each has a no-terminal desktop app.

Round 2: Codex vs Claude Code pricing

Prices below are for one person, as of November 2026, in US dollars.

Plan levelCodex (inside ChatGPT)Claude Code (inside Claude)
FreeYes, in the desktop app (subject to rollout)Not included
EntryGo, $8/month, desktop appNone below Pro
StandardPlus, $20/month: web, CLI, IDE extension, iOSPro, $20/month or $17/month billed yearly
Heavy usePro at $100, $200 or $500 a monthMax $100/month (5x Pro usage), Ultra $200/month
TeamsBusiness, $20/user/month yearly or $25 monthlyTeam, $20/seat/month yearly or $25 monthly
Pay per useOpenAI API keyAnthropic API key

At $20 a month they cost the same. Below $20, only Codex has an option, and the free tier lets you see what an agent does before you spend anything. Be aware that the Free and Go plans are listed for the desktop app only. The terminal tool and the editor extension are listed from Plus upward.

At the top end both companies sell $100 and $200 plans, and OpenAI adds a $500 tier. What you really buy at that level is how long you can work before you hit a usage limit. Those limits are not printed as fixed numbers and they shift when new models arrive, so compare them on the official pages (Claude pricing, Codex pricing). Our Claude Code pricing guide breaks down what each Claude plan gets you.

Round goes to: Codex. A free way in and an $8 plan beat no free option at all.

Round 3: how each agent works with your files

Both agents work the same basic way. You type a request in plain English. The agent reads the files it thinks matter, makes a plan, edits files, runs commands such as tests, reads the output and tries again if something failed. You see each change as a diff (a before and after view where removed lines and added lines are marked).

Claude Code leans hard on planning. Its plan mode lets it look around your project and write a plan without changing a single file. You read the plan, ask for changes, and only then let it start. That is useful on bigger jobs where a wrong first step wastes time. Read more in our plan mode guide. Claude Code can also split a job between subagents (helper agents that each handle one part, like research or testing) so the main chat stays tidy.

Codex keeps things simpler by default. It works in the folder you started it in, and on paid plans it can also run tasks in the cloud. A cloud task works on a copy of your code on OpenAI's computers, so you can close your laptop and check the result later on the web or your phone.

If you want the agent to work on a separate copy of your project so your main folder stays clean, both can do it with a git worktree (a second copy of your project folder that git keeps linked to the first). Each agent then edits its own copy, and you merge the parts you like back into the main folder.

Round goes to: Claude Code. Plan mode and subagents give you more say over how a big change happens.

Round 4: safety, approvals and permission modes

This round is about one question: what can the agent do without asking you first? Both tools ask before risky actions by default, and both let you loosen that once you trust them.

Claude Code's permission modes

  • Manual (the default): reads files freely, but asks before it edits a file or runs most commands.
  • Accept edits: edits files and runs simple file commands (like making a folder) without asking. It still asks before other commands.
  • Plan: reads and explores. Nothing is edited until you approve its plan.
  • Auto: skips routine prompts, while a background checker reviews each action before it runs.
  • Bypass permissions: runs everything without asking. Anthropic says to use it only inside an isolated container or virtual machine (a fenced-off test computer).

In the terminal you press Shift+Tab to move between modes. When you answer a prompt with "Yes, and don't ask again" for a command, Claude Code saves that rule in a local settings file (.claude/settings.local.json), and that project stops asking about it. There is also a sixth mode, dontAsk, which quietly refuses anything that would need a prompt. You can only switch it on from the command line.

Codex's sandbox and approvals

Codex wraps its commands in a sandbox (a fenced-off area the agent cannot write outside). In a folder that uses git, it starts in Auto, which means it can write inside your project and asks you before going further. In a folder without git it starts read-only. Network access is off by default, so the agent cannot download things or call websites unless you turn that on. You can drop to read-only at any time with the /permissions command.

Codex's starting point is stricter, because network access stays off until you allow it. Claude Code gives you more modes and finer rules, plus hooks (small scripts you write that run before or after the agent acts) that can block a command before it runs.

Round goes to: Claude Code, narrowly. More ways to set exactly how much freedom the agent gets, though Codex's stricter starting point is a fair argument for the other side.

Round 5: project instructions (CLAUDE.md and AGENTS.md)

Every agent forgets your preferences when a new chat starts. A project notes file fixes that. You write your rules once ("use Tailwind", "never touch the payments folder", "run the tests before you say you're done") and the agent reads them at the start of every session.

Codex reads AGENTS.md. Claude Code reads CLAUDE.md, and if a project has no CLAUDE.md, recent versions read AGENTS.md instead. Claude Code goes further with notes at several levels: one for you across all projects, one for the project, and smaller ones inside sub-folders that load only when the agent works there. It also adds skills (saved instructions for a repeated job) and MCP connections that let the agent read outside tools like your issue tracker.

Tip

If you keep both agents on one project, put shared rules in AGENTS.md. Then start your CLAUDE.md with the line @AGENTS.md and add Claude-only extras below it. Once a CLAUDE.md exists, Claude Code stops reading AGENTS.md on its own, and that line pulls it back in. In VibeCodingDeck you can open and edit these notes files for each project from the app, so you never hunt for them in a folder.

Round goes to: Claude Code. Layered notes, skills and MCP give it more places to store how you like to work.

Round 6: how fast everyday work feels

We are not going to quote speed tests here. No official, fair speed comparison exists for these two tools, and any number would be out of date after the next model release. What we can describe is how daily use feels.

Codex tends to feel quick on short, clear jobs: fix this error, rename this, add a button. It gets going without much back and forth. Claude Code often spends more time reading your project and planning before it edits. On a small job that can feel slow. On a big job that touches many files, the planning often saves you a second or third attempt.

A bigger delay than thinking speed is running into your plan's usage limit halfway through a task. Then you wait for the limit to reset, whichever tool you use. If you read codex vs claude code reddit threads, check the date on each post first. A new model release can flip people's opinions, so a thread from six months ago may describe tools that have since changed. Read those posts as one person's experience, not as test results.

Round goes to: Codex. It gets small jobs done with less waiting, and small jobs are most of a normal day.

Round 7: Codex and Claude Code on Windows

Both run on Windows without WSL (a Linux system inside Windows). Claude Code installs from PowerShell with the command in round 1, and Anthropic recommends having Git for Windows installed. The Claude desktop app has Windows builds for normal x64 computers and for ARM64 laptops.

Codex runs natively on Windows too, and its sandbox works there, so the agent stays fenced inside your project folder. Claude Code is different here. Anthropic's setup page says its command sandbox is not supported on plain Windows and needs WSL 2. The Codex desktop app is available from the Microsoft Store.

Common snags are the same on both: PowerShell blocking scripts, the tool not found after install until you open a new terminal, and antivirus slowing the first run. Step-by-step fixes are on our Claude Code on Windows and Codex on Windows pages.

Round goes to: Codex. Both install easily, but only Codex keeps its sandbox on plain Windows.

Which is better, Codex or Claude Code, for you

The score ended 3 rounds each, with one tie. Codex took price, everyday speed and Windows. Claude Code took file handling, safety settings and project notes. That leaves your budget and the size of your jobs to break the tie.

Codex

you want to try an AI coding agent for free, or you already pay for ChatGPT.

Codex

your budget is under $20 a month and your tasks are mostly small fixes.

Claude Code

you build bigger features and want the agent to plan before it edits.

Claude Code

you want fine control over what the agent may do, with hooks and permission rules.

Both

you code most days and can't afford to stop when one plan hits its limit.

Codex vs Claude Code for beginners

If you have never used an AI coding agent, start with whichever company you already pay. If you pay for neither, Codex's free desktop app is the cheapest way to learn what an agent does. Once you outgrow it, Claude Code's plan mode is very friendly for beginners because you read a plain-English plan before anything changes. Pick a tiny first project, such as a one-page site about a hobby, so mistakes cost nothing.

Running both at once is easier than it sounds. Our page on running multiple agents shows how to keep them in separate project copies so they never edit the same file at the same moment.

Codex and Claude Code: quick answers

Is Codex better than Claude Code?

+

Not across the board. In our rounds Codex won on price, quick everyday jobs and Windows. Claude Code won on planning, safety settings and project notes. The lead can change with each new model, so check again after big releases.

Is OpenAI Codex free?

+

As of November 2026, the ChatGPT Free plan includes Codex in the desktop app, subject to rollout. The terminal tool and the editor extension are listed from the $20 Plus plan upward, or you can use an OpenAI API key.

Can I use Claude Code for free?

+

No. The free Claude plan does not include Claude Code. You need Pro ($20 a month, or $17 a month billed yearly as of November 2026), Max, Ultra, a Team seat or an Anthropic API key.

Can Codex and Claude Code work on the same project?

+

Yes. Put shared rules in AGENTS.md, and start your CLAUDE.md with the line @AGENTS.md so Claude Code reads them too. To stop them from overwriting each other, give each one its own worktree (a separate linked copy of the project) and merge the good work back.

Which models do Codex and Claude Code use?

+

Codex runs OpenAI's GPT models; OpenAI's pricing page lists GPT-6 Luna for the free desktop tier as of November 2026. Claude Code runs Anthropic's Claude models. Neither lets you swap in the other company's models.

Is the Codex CLI open source?

+

Yes. The Codex command-line tool is published at github.com/openai/codex under the Apache-2.0 licence. Claude Code's command-line tool is not open source.

Sources

Run up to 12 AI coding agents in one window.

Claude Code, Codex, Gemini, Grok and more, side by side. Windows and Mac. From $29 a month.

Get started

Keep reading