Codex vs Claude Code: Plans, Sandboxing and Fit

Codex is OpenAI's coding agent and Claude Code is Anthropic's. Both read your repository, edit files, run commands and open pull requests. The useful differences sit in the plumbing: which plan you need, what the agent may touch before it asks, where instructions live and how each connects to GitHub. No benchmarks here. Your codebase is the only test that counts.

The short answer

• Choose Codex if your team runs on ChatGPT, wants a sandbox on native Windows, keeps repositories on GitLab as well as GitHub, or wants safe defaults with no setup.
• Choose Claude Code if you prefer Claude models, work in JetBrains or on Android, or want to tune permissions per tool and command.
• Run both if you can. One AGENTS.md can steer both, and both extensions install in VS Code.

Codex vs Claude Code at a glance

Pricing and plans

Codex pricing (as of November 2026)

Codex comes with ChatGPT. OpenAI's Codex pricing page lists Free ($0) and Go ($8 a month) with GPT-6 Luna at standard speed in the desktop app, subject to rollout. Plus ($20 a month) adds Codex on the web, in the CLI, in the IDE extension and on iOS, cloud integrations such as automatic code review and Slack, GPT-6.1 Sol and GPT-6 Luna, and the option to buy credits when you run out. Pro starts at $100 a month, with tiers at $100, $200 and $500; the $500 tier adds Ultrafast mode, and OpenAI says Pro plans currently have no five-hour limit.

Business is $20 per user a month billed annually (two users minimum) or $25 monthly, with SAML SSO. Enterprise and Edu go through sales. Signing in with an API key instead means API rates and the CLI, SDK and IDE extension only, without cloud features such as GitHub review or Slack.

Claude Code pricing (as of November 2026)

Anthropic's pricing page marks Claude Code as included on Pro, Max and Ultra, and not on Free. Pro is $20 a month, or $17 a month with $200 billed up front. Max is $100 a month with five times Pro usage every five hours and every week. Ultra is $200 a month with ten times Pro weekly usage and no five-hour limit for verified accounts. Fable, which Anthropic's docs call the most capable models in Claude Code, is paid with usage credits on Pro and included up to half the weekly limit on Max.

Team (2 to 150 people) has Standard seats at $20 per seat a month billed annually or $25 monthly, and Premium seats at $100 or $125 with five times the usage. Enterprise is US$20 per seat a month billed annually, with a 20-seat minimum, plus usage at API rates. A Claude Console account gives pay-per-token billing instead. One thing to know before you compare limits: Codex shares its allowance with ChatGPT Work, and on Claude Team and Enterprise seats Claude Code shares one with Claude chat and Cowork. Long non-coding sessions eat into your coding time.

Where you can run each

The open-source Codex CLI installs with a script, npm or Homebrew. The IDE extension installs in VS Code, Cursor and Windsurf; Xcode and JetBrains offer Codex through their own integrations. The ChatGPT desktop app runs each Codex chat locally, in a Git worktree or in the cloud. Codex Cloud gives each task its own workspace that keeps going while your laptop sleeps, and the ChatGPT iOS app can start cloud tasks or approve commands on a connected Mac or Windows PC.

Claude Code runs in similar places. The terminal CLI is the fullest version (scripting and the Agent SDK are CLI-only). There's a VS Code extension that also installs in Cursor, a JetBrains plugin that needs the CLI, a desktop app on macOS and Windows (Linux in beta), the web at claude.ai/code, and the Claude app for iOS and Android. Remote Control lets you drive a session on your own machine from the browser or your phone.

Models

Codex runs OpenAI models. Its docs currently recommend GPT-6 Astra for the most demanding work, GPT-6.1 Sol as a near-Astra option at lower cost, GPT-6.1 Terra for fast and efficient work, and GPT-6 Luna for focused, repeatable tasks. You switch with /model or codex --model. With an API key, your choice follows the models your key can reach.

Claude Code runs Claude models through aliases: opus, sonnet and haiku resolve to Opus 5.5, Sonnet 5.5 and Haiku 5.5 on the Anthropic API, and fable resolves to Fable 5.2, which no plan uses as its default. The opusplan alias plans with Opus and executes with Sonnet. Claude Code also works through Amazon Bedrock, Google Cloud and Microsoft Foundry, where some aliases point at older versions.

Sandboxing and approvals

This is the biggest practical difference. Codex sandboxes local commands from the first run, enforced by the operating system (Seatbelt on macOS, bubblewrap on Linux and WSL2, a native sandbox on Windows): network off, writes limited to your workspace. An approval policy decides when Codex stops to ask, such as before editing outside the workspace or using the network. /permissions switches presets, including read-only.

Claude Code starts from permission modes. Manual mode asks before most edits, shell commands and network calls; acceptEdits lets file edits through; plan blocks edits until you approve a plan; and auto mode, the starting mode for terminal and VS Code sessions from version 2.1.283, has a second model review each action instead of you. Its Bash sandbox, which limits what shell commands can write and which hosts they reach, is off by default, so run /sandbox to switch it on. It works on macOS, Linux and WSL2; on native Windows commands run unsandboxed. If you want strict defaults without configuring anything, Codex is ahead here. If you want to allow npm test and block git push by rule, Claude Code gives you finer controls.

Instruction files: AGENTS.md vs CLAUDE.md

Codex reads AGENTS.md. It loads a global file from ~/.codex, then walks from the repository root to your working directory, joining every AGENTS.md it finds so deeper files override earlier ones, up to 32 KiB by default. AGENTS.override.md replaces a level temporarily, and /init writes a starter file.

Claude Code reads CLAUDE.md from your home folder (~/.claude/CLAUDE.md) and from the project (./CLAUDE.md or ./.claude/CLAUDE.md), plus a git-ignored CLAUDE.local.md and topic files in .claude/rules/. It also keeps an auto memory of your corrections. When a repository has an AGENTS.md and no CLAUDE.md, Claude Code reads the AGENTS.md. The simplest way to serve both agents is one AGENTS.md, imported from a short CLAUDE.md if you need Claude-only notes. What to put in it hardly changes between tools; our best practices for AI coding cover the habits that matter.

MCP, skills and hooks

Both agents speak the Model Context Protocol, so the same servers work in each. In Codex you add one with codex mcp add, for a local command or a remote URL with OAuth; in Claude Code it's claude mcp add --transport http for a remote server. If MCP is new to you, start with our explainer on what MCP is and why it matters.

The extension systems have converged. Both support skills, plugins, hooks that run scripts at points in the agent's lifecycle, and subagents that take a focused piece of work and report back. For scripts and CI, Codex has codex exec and the Codex SDK; Claude Code has claude -p and the Agent SDK.

GitHub and code review

Codex Cloud checks out repositories from your connected GitHub or GitLab account. For review, you comment @codex review on a GitHub pull request or turn on automatic reviews, and Codex posts a normal review that flags only P0 and P1 issues, following any ## Code Review Rules section in your AGENTS.md. GitLab merge requests get the same @codex review flow. For CI there's the openai/codex-action@v1 GitHub Action.

Claude Code's cloud sessions clone from GitHub through the Claude GitHub App or your local gh token; GitLab and Bitbucket repositories have restrictions on the web. Its GitHub Action answers @claude in pull requests and issues, and /install-github-app sets it up for github.com repositories. The managed Code Review service, where several agents review each pull request and verify findings before commenting, is for Team and Enterprise plans. On any plan, /code-review runs a review inside your session.

Who each is for

• Codex fits ChatGPT users, Windows developers who skip WSL, GitLab shops, and anyone who wants an agent locked down by default.
• Claude Code fits Claude users, JetBrains and Android users, and teams that want permission rules per command.

Limits of each

• Codex: API-key sign-in drops every cloud feature, including GitHub review and Slack, so CI-only setups lose a lot.
• Codex: usage on Plus is an estimate per five-hour window that varies by model, and weekly limits may also apply.
• Codex: the IDE extension and Codex Cloud are not open source; only the CLI, SDK and some components are.
• Claude Code: no free tier, and Fable models cost usage credits on Pro.
• Claude Code: the Bash sandbox is off until you enable it, and native Windows has no sandbox at all.
• Claude Code: cloud sessions and the managed review service center on GitHub.

Which should you choose

Start with the subscription you already have: ChatGPT Plus includes Codex and Claude Pro includes Claude Code, so try the one you already pay for on a real task. If you pay for neither, pick by environment: native Windows or GitLab points to Codex; JetBrains, or a wish to approve commands rule by rule, points to Claude Code. Then write one AGENTS.md, give both agents the same bug, and compare the diffs. You'll know in an afternoon. If you're also weighing an editor-first tool, our Cursor vs Claude Code comparison covers that trade.

Alternatives worth a look

If neither agent fits, our list of AI pair programming tools covers GitHub Copilot, Cursor and others with current plans, and our roundup of open-source AI coding tools is the place to start if you want to self-host or bring your own model.

Frequently Asked Questions

Is Codex or Claude Code cheaper?
At the entry level they cost the same: ChatGPT Plus is $20 a month and includes Codex, and Claude Pro is $20 a month (or $17 a month with $200 billed up front) and includes Claude Code. Codex also has limited access on ChatGPT Free and Go ($8 a month); Claude Code is not on Claude Free. Above that, ChatGPT Pro comes at $100, $200 or $500 a month and Claude Max is $100 and Ultra $200. Prices as of November 2026.

Can I use Codex or Claude Code for free?
Codex, yes, with limits: OpenAI lists limited Codex access on ChatGPT Free, and its Codex pricing page says Free runs GPT-6 Luna in the desktop app, subject to rollout. Claude Code, no: Anthropic's pricing page marks it as not included on Free. Both can also run on pay-as-you-go API billing (an OpenAI API key for Codex, a Claude Console account for Claude Code).

Does Claude Code read AGENTS.md?
Yes. Anthropic's docs say Claude Code reads AGENTS.md as your project instructions when there is no CLAUDE.md or CLAUDE.local.md in the working directory or above it. If both exist, it reads CLAUDE.md only, unless that file imports AGENTS.md or you change the Project instructions setting. Codex reads AGENTS.md natively, plus a global file in ~/.codex.

Which one is safer to run on my laptop?
Codex sandboxes local commands by default, with no network access and writes limited to the workspace, on macOS, Linux, WSL2 and native Windows. Claude Code relies on permission modes by default; its Bash sandbox is off until you run /sandbox, and it works on macOS, Linux and WSL2 only.

Can I use both in the same editor?
Yes. The Codex extension installs in VS Code, Cursor and Windsurf, and the Claude Code extension in VS Code and Cursor. Each needs its own plan or API billing.

Do Codex and Claude Code review pull requests?
Both can. Codex reviews GitHub pull requests when you comment @codex review or turn on automatic reviews, and its Plus plan lists automatic code review. Claude Code answers @claude mentions through its GitHub Action, authenticated with a Claude API key or a token from a Pro, Max, Ultra, Team or Enterprise plan, and /code-review runs a local review on any plan. Anthropic's managed Code Review service is for Team and Enterprise plans on GitHub.