Claude Code vs Codex vs Copilot comes down to three sentences. Claude Code is Anthropic's terminal agent: it reads your files, edits them and runs commands, without GitHub and without any programming knowledge. Codex is OpenAI's counterpart, per its docs also a terminal agent with an editor extension and a cloud variant, included in every ChatGPT plan. GitHub Copilot is the one most tied to GitHub: editor completions, a cloud agent that produces pull requests, and its own command line. If you want to work on files and folders in your own business, Claude Code or Codex fits; if you already live in GitHub, Copilot gives you the shortest path to a pull request.
I have used Claude Code every day since it came out, and I run my AI workforce on it. I use Codex regularly, to see how it changes and for correction loops. GitHub Copilot I tested, but never ran day to day. Prices and feature lists still come from the vendors' official docs, as of September 2026; where my own impression comes in, I say so. Everything else I write about Claude is collected on the Claude page.
Three tools, three basic ideas
The three tools sound alike because all three are called "agents" and all three can run in a terminal. The difference lies in where they come from. According to Anthropic's documentation, Claude Code is "an agentic coding tool that reads your codebase, edits files, runs commands", available in the terminal, in VS Code, Cursor and JetBrains, as a desktop app for macOS and Windows (Linux in beta), on the web at claude.ai/code including the mobile app, in GitHub Actions, GitLab CI/CD, Slack and Chrome. It requires a paid Claude plan or an Anthropic Console account.
OpenAI describes Codex in its CLI documentation like this: "Inspect code, make changes, run commands, and automate repeatable work without leaving your terminal." There is the command line, an extension for VS Code, Cursor and Windsurf, a cloud variant in the browser, a desktop app for macOS and Windows, and a GitHub integration. You sign in with your ChatGPT account or an API key, and you set permissions with /permissions. In its basic idea, Codex is the closest to Claude Code: an agent that works in your folder.
GitHub Copilot is built differently. Its core is completion in the editor; on top of that come chat, a cloud agent and a command line called Copilot CLI, according to the GitHub docs. Per the vendor, the cloud agent runs in its own short-lived environment on GitHub Actions, at most 59 minutes per session, only for GitHub repositories, and delivers exactly one pull request per task. Copilot is less a tool for your folder than a tool for your repository.
Copilot vs Claude Code vs Codex: the comparison table
The table sums up what the vendors state in their docs (as of September 16, 2026). Prices are entry prices in US dollars; the full pricing logic follows below.
Two things stand out. First, the free entry points: two of the three have one, Claude Code does not, it requires a paid plan per Anthropic. They are only partly comparable, though. GitHub states an allowance in numbers for Copilot Free, 2,000 completions a month plus Copilot CLI. Codex is included in the free ChatGPT plan per OpenAI's docs, but OpenAI states no comparable allowance in numbers for it; all that is documented is that the allowance grows with the paid tiers. So if you try Codex for free, you do not know in advance how far you get. Second: Copilot brings in both competitors. GitHub's plan overview lists "Access to 3rd party agents (Claude Code and Codex)" for Pro. So you do not necessarily have to choose between the three if your team already pays for Copilot.
Prices and usage logic compared
I have summarized Claude's prices with sources in Claude Pricing and Plans Explained, as of September 2026: Free 0 dollars (without Claude Code), Pro 20 dollars a month, Max 100 or 200 dollars a month, Team from 20 dollars per seat a month. Anthropic counts usage in five-hour windows according to its pricing page; Max offers five or twenty times the usage per window that Pro does.
OpenAI's Codex pricing docs (as of September 16, 2026) list Free, Go at 8 dollars and Plus at 20 dollars a month, with Codex included in all three; Pro starts at 100 dollars a month, with five or twenty times higher rate limits than Plus; Business costs 25 dollars per user monthly or 20 on annual billing. OpenAI also counts messages per five-hour window, plus weekly limits and purchasable credits.
GitHub counts differently according to its own docs: Free at 0 dollars with 2,000 completions a month and Copilot CLI, Pro 10 dollars, Pro+ 39, Max 100 a month, Business 19 and Enterprise 39 per seat and month. Agent usage is billed in "AI credits", per the docs 1,000 for Pro, 3,900 for Pro+ and 10,000 for Max.
What does that mean for the comparison? Three vendors, three units of measure: five-hour windows, messages plus a weekly limit, credits. A direct cost comparison is therefore only a rough heuristic. From my own experience with Claude Code: I did not start on Max, I moved only when Pro's allowance stopped covering long sessions. For business data I also use Claude Team, because of the data processing agreement and privacy.
What makes the difference if you do not code
If you do not write software but want to work on proposals, meeting notes, checklists and folder structures, three questions count: Does the tool run in a normal folder without a repository? Can I set rules that always apply? Can I reuse workflows?
Claude Code answers all three with yes, shown step by step in Claude Code: A Guide for Beginners. A CLAUDE.md in the project folder is the house rule, skills are reusable workflows, and hooks are rules that run on every event; more on that in Claude Code Hooks: Rules That Always Apply. Codex also works in any folder with permission levels. I use it regularly, but for correction loops rather than as the home of my workforce, so how far rules and reusable workflows carry there I have not tested systematically. Copilot requires Node 22 and a GitHub account for its command line, and the cloud agent needs a GitHub repository. For a business without GitHub, that is a high hurdle.
An example from my workforce that uses exactly these three points: Sebastian, my SEO employee, keeps his workflows as skills named seo-audit and geo-audit in his folder. On September 16, 2026 he asked Conny, my content editor-in-chief, through our mailbox protocol for her experiences for five blog posts; Conny answered as a subagent and dropped the letter into Sebastian's inbox. No repository, no pull request, just a folder structure with rules in which two Claude Code sessions work together.
My choice: Claude Code, and where its limits are
My choice was made before the other two existed in their current form: I have used Claude Code every day since its start as a research preview on February 24, 2025, and my AI workforce runs on it. Each has its own CLAUDE.md as a personnel file, its own skills and its own mailbox, and Conny and Ralf regularly run in long, continuous sessions. My most concrete hook use case: six hook events report to my employee OS, and when a permission request comes up I can decide there instead of in the terminal; that has been running live, verifiably, since September 2026.
To be honest, there are limits. Claude Code has no free tier, and with heavy use you hit the five-hour windows. A point-and-click interface is not the concept, even though a desktop app and a web version exist. And if your team is already organized on GitHub, Copilot gives you a pull request workflow that I would have to build myself with Claude Code. How Claude Code differs from the desktop tool Cowork is covered in Claude Code or Cowork: What For?.
One tool writes, another checks
One house rule applies here: whoever creates something does not correct it. The reasoning is the same as for the review pass in my workforce, that a model is poor at finding its own mistakes in the same context. In practice one tool writes and another reads it back, and that is what I use Codex for. Claude Code builds something, Codex checks the result with a fresh view, or the other way around.
This is not a benchmark, I measure nothing here. What counts is the separation, not the vendor: two separate sessions of the same tool work too, as long as the checker does not know the path to the result.
When Codex or Copilot is the better choice
My assessment, for Codex from ongoing use, for Copilot from the docs and one test:
- You already pay for ChatGPT Plus and do not want a second vendor: Codex is included, and the basic idea is the same as Claude Code's. Whether the model behind it suits you better is a separate question; for chat and text work I answered it in Claude vs ChatGPT 2026: Compared.
- Your team lives in GitHub, every change goes through pull requests, and you want to hand tasks to an agent in a clean environment: that is what the Copilot cloud agent is for, included in all paid plans per the GitHub docs.
- You want to try things out for free first: Copilot Free and Codex in the free ChatGPT plan both work without a subscription, Claude Code does not. Only GitHub names an allowance in numbers, 2,000 completions a month plus Copilot CLI; for Codex on the free plan OpenAI names none.
- You want Claude Code but your company only has Copilot: according to GitHub's plan overview, Copilot Pro includes access to Claude Code and Codex as third-party agents, for the paid individual plans and not for Free. The page does not state which limits apply, so check that in your own account.
If none of these four cases applies and you want to work with files in your own business, Claude Code is in my view the choice that needs the least infrastructure.
Frequently asked questions
Can I use Claude Code inside GitHub Copilot?
According to GitHub's plan overview (as of September 16, 2026), Copilot Pro includes "Access to 3rd party agents (Claude Code and Codex)". What that means in allowance is not stated. If you want Claude Code as a standalone tool in your terminal, you still need a Claude plan.
Do I need a paid ChatGPT plan for Codex?
No, according to the Codex pricing docs it is also included in the free ChatGPT plan. The allowance grows with Go, Plus and Pro. For regular work in the terminal, a paid tier is realistic, just as with Claude.
Claude Code vs Copilot: which one runs without GitHub?
Claude Code and Codex run in any folder, no repository required. Copilot is tied to a GitHub account, and the cloud agent to a GitHub repository. Copilot CLI runs locally but also needs the GitHub account.
What is the difference between a cloud agent and the terminal?
In the terminal, the agent works directly on your machine, in your files, and asks before risky steps. A cloud agent like Copilot's or Codex's works in an environment at the vendor, fetches your repository and returns a pull request. For folders without a repository, the terminal is the only way.
Claude Code vs Codex 2026: which tool is better if you do not code?
Both run in a normal folder and without a repository according to their docs. For Claude Code I can say from my own use that CLAUDE.md, skills and hooks hold up for pure text work as well. I use Codex regularly, but for correction loops rather than as a place to keep rules and workflows, so I have not tested it for that purpose.
Does any of the three tools run on Windows without WSL?
According to the docs, yes, all three: Claude Code via the PowerShell installer or WSL, Codex "native and WSL", Copilot CLI via PowerShell and WSL. The Windows steps are in Install Claude Code: Windows, Mac, Linux.
Where to go next
If you decide on Claude Code, start with Claude Code: A Guide for Beginners: installation, CLAUDE.md and your first task. If you are torn between terminal and desktop, Claude Code or Cowork: What For? helps. And if your real question is whether Claude or ChatGPT suits you better, Claude vs ChatGPT 2026: Compared is the starting point. How my AI employees are built on Claude Code is what I show in my community.
