the agent office
Codex vs Claude Code (2026): which to use, and how to run both
Last updated October 12, 2026. Models, plans and prices change often, so check each vendor's page before you buy.
Codex and Claude Code are the two coding agents most developers compare right now. Both read your repo, edit files, run your tests and open a pull request. Both are good. Which one suits you depends on how you work, what you already pay for and what kind of tasks fill your week.
The short answer
Pick Claude Code if you have no OpenAI subscription yet, you like to steer an agent as it works, or you want one tool that covers terminal, editor, desktop app and browser with the same settings and project instructions everywhere.
Pick Codex if you already pay for ChatGPT, you want a free way to try an agent, or you want its command line tool, editor extension and cloud version tied to one ChatGPT login.
Pick both if you are not sure. We cannot tell you which one writes better code for your repo. The honest way to find out is to give each one a real ticket and read both pull requests. The last section shows how.
Side by side
| Claude Code | Codex | |
|---|---|---|
| Made by | Anthropic | OpenAI |
| Where it runs | Terminal, VS Code and JetBrains extensions, desktop app, web [1] | Command line, IDE extension, ChatGPT desktop and mobile apps, ChatGPT on the web, Codex Cloud [4] |
| Models | Defaults to Opus 5.5 on Pro, Max, Team, Enterprise and the Anthropic API. Aliases for Sonnet, Opus and Fable [2] | GPT-6.1 Sol and GPT-6 Luna on Plus and above [3] |
| Free tier | No. Claude's free plan does not include Claude Code [5] | Yes, with limits. Free and Go plans get GPT-6 Luna in the desktop app [3] |
| Entry paid plan | Pro, $20 a month ($17 on annual billing) [5] | Plus, $20 a month [3] |
| Heavier plans | Max from $100 a month, with 5x or 20x Pro usage [5] | Pro at $100, $200 or $500 a month [3] |
| Team plans | Team Standard seat $20 a month annual, Premium $100 annual [5] | Business, $20 per user a month, minimum 2 users, annual billing [3] |
| Pay per use | Anthropic API [1] | API key, billed at standard API rates [3] |
| Project instructions | CLAUDE.md. Can read an existing AGENTS.md in its place [1] |
Not checked for this post |
| Extras | Subagents, hooks, MCP, skills, scheduled tasks, GitHub Actions [1] | Sandboxing, approval settings, auto-review, code review and Slack integrations on Plus [3][4] |
| CLI source | Not stated in the docs we read | Open source, Apache-2.0 [6] |
Sources, all read on October 12, 2026:
- Claude Code overview
- Claude Code model configuration
- Codex pricing
- Codex documentation
- Claude plans and pricing
- openai/codex on GitHub, license shown in the repository's metadata
Two cautions about the table. First, the prices are for subscriptions, and the usage you get for the money is not the same on both sides. Both vendors say allowances vary with the model you pick and the size of the task, so a $20 plan on one is not the same amount of work as a $20 plan on the other. We have not measured that. Second, a model name tells you little about how it will do on your code. Treat the model row as a fact about what you get, not a score.
Where each one wins
Here is the part that is easy to overclaim, so we will be plain about it. We have not run a controlled benchmark of these two tools for this post. What follows is general guidance based on how each vendor positions its tool and how developers tend to use them. Treat it as a place to start testing, not a measured result.
Where Claude Code is a good fit
A change that spans many files. Say you want to rename a concept across the app: the database column, the API, the screens and the tests. Claude Code's documentation describes it as planning an approach and then writing code across multiple files, and it is built to be steered along the way. If you want to read the plan and say "no, not that file," it suits that style.
Work that depends on your own setup. Claude Code's hooks, skills and MCP connections let you wire in your own checks, such as running a formatter after every edit or reading tickets from Jira. If your team already has that set up, the agent picks it up on every surface.
Scheduled and background jobs. The docs list routines that run in the cloud, desktop scheduled tasks and a loop command, which suit chores like a morning review of open pull requests.
Where Codex is a good fit
Clearly written tasks you want many of at once. Say you have five small tickets, each with a clear repro and a clear definition of done. Codex has a cloud version [4], and running several tasks there at once is the kind of work it is commonly used for. The pricing page lists web access starting at Plus [3], so check your plan first.
A low-cost first try. If you want to see what an agent feels like without paying anything, Codex has a free plan and Claude Code does not. For a first taste that can matter more than anything in the table.
An open source command line tool. If you want to read or build on the code of the thing running on your machine, the Codex CLI is Apache-2.0 licensed on GitHub.
Staying inside one account. If your team already lives in ChatGPT, Business seats bring Codex along, with workspace controls such as SSO.
Where it is a tie
Both can open pull requests, both ask before risky actions, both work with your editor. The differences you notice day to day are likely to be about taste: how the agent talks, how much it asks before acting and how it handles your particular codebase. That is exactly why testing beats reading.
Run both on real tickets
You do not need to choose before you try. Here is a simple experiment that takes an afternoon.
- Pick two similar tickets. Choose two small, real tickets from your backlog that are about the same size. A bug with a clear repro and a small feature both work. Better still, use the same ticket twice, if you have time to throw one result away.
- Give each one its own worktree. A git worktree is a separate copy of your repo on its own branch. It means two agents never edit the same files. Setting this up by hand for every ticket is the tedious part.
- Put Codex on one and Claude Code on the other. Same instructions, same definition of done.
- Compare the pull requests. Read both. Which one has the better tests? Which one changed less than it needed to, or more? Which needed less nudging from you? Which would you rather merge?
The Agent Office does steps 2 and 3 for you. It is a small office on your Mac. Connect Linear, Jira, GitHub Issues or 8090, and your tickets show up on a board. Hire a worker for a ticket, choose Codex or Claude Code for it, and the worker gets its own worktree and branch. Anything risky, like pushing code or installing a package, waits for your OK. When both finish, you have two pull requests side by side.
Run the experiment on three or four tickets, not one. A single result can come down to luck, while a pattern across a few tickets tells you something about your code and your style.
It also helps with cost. Once you see which tool handles which kind of ticket well, you can send the routine ones to the cheaper or free option and save the heavier plan for the hard ones. For a wider look at the field, see our guide to the best AI coding tools in 2026, and for the setup in more depth, how to run multiple AI coding agents.
How to choose
- You already pay for ChatGPT: start with Codex, since it is already in your plan.
- You already pay for Claude: start with Claude Code, for the same reason.
- You pay for neither and want to spend nothing: Codex on a free ChatGPT account is the way in.
- You are paying for one and wondering if the other is better: run the two-ticket experiment before you add a second bill.
Whatever you choose, expect this page to go out of date. Both vendors ship new models and change their plans every few months. We will update the date at the top when we recheck.