Codex vs Claude Code: which coding agent to use

Codex (OpenAI) and Claude Code (Anthropic) are the two most used terminal coding agents, and they now overlap almost feature for feature: a CLI, an IDE extension, a desktop app, cloud tasks, hooks, MCP and plugins. The real differences are the models, how each one sandboxes commands by default, whether you can read the source, and which subscription you already pay for. If you have ChatGPT Plus, start with Codex; if you have a Claude plan, start with Claude Code; many developers end up running both.

As of October 2026. Both tools ship weekly, so check the linked docs before you commit.

Summary

What each one is

Codex is OpenAI’s coding agent. The Codex CLI is open source under Apache-2.0 and installs with npm install -g @openai/codex, brew install --cask codex or a one-line script. The same agent runs in the IDE extension, in Codex Cloud on the web and on iOS, and in the ChatGPT desktop app, where a switch picks ChatGPT or Codex. Running codex app opens that desktop experience, which is what people search for as the “Codex app” or “Codex desktop”.

Claude Code is Anthropic’s coding agent. It installs with curl -fsSL https://claude.ai/install.sh | bash (or Homebrew, WinGet, apt), then you run claude in a project. It also runs in the VS Code extension (which installs into Cursor too), a JetBrains plugin, the Claude desktop app, and on the web and the Claude mobile app. Its docs note that every surface shares the same engine, so your CLAUDE.md, settings and MCP servers follow you.

Models

Codex runs OpenAI models. The models page recommends GPT-6.1 Sol for complex coding and agentic work, GPT-6 Astra for the hardest end-to-end tasks, and GPT-6 Luna for focused, high-volume jobs. With an API key you can use any API model your key has.

Claude Code runs Claude models. On Pro, Max, Team, Enterprise and the API it defaults to Opus 5.5. /model switches to Sonnet 5.5, Haiku 5.5 or Fable 5.1, and current models run with a 1M-token context window. It can also run through Amazon Bedrock, Google Cloud, Microsoft Foundry or your own LLM gateway, which matters if your company buys AI through a cloud account.

Neither runs the other company’s models. If you want one tool for many model vendors, see OpenCode vs Claude Code.

Pricing and limits

Codex (pricing):

• Free and Go ($8/month): GPT-6 Luna in the desktop app, still rolling out.
• Plus ($20/month): Codex on the web, CLI, IDE extension and iOS, with GPT-6.1 Sol and Luna. OpenAI estimates 15–160 local GPT-6.1 Sol messages per five hours, plus weekly limits.
• Pro ($100, $200 or $500/month): no five-hour limit for now. Ultrafast mode is on the $500 tier.
• Business: $20 per user per month billed yearly, $25 monthly.
• API key: pay per token. Works in the CLI and IDE extension, not cloud tasks.

Cloud tasks draw from the same allowance as local messages and can use more of it. Plus and Pro can buy credits after a limit.

Claude Code (pricing):

• Free: no Claude Code.
• Pro: $20/month, or $17/month billed yearly.
• Max: from $100/month, with 5x or 20x Pro’s usage.
• Team: Standard seats $25/month ($20 yearly), Premium $125 ($100 yearly).
• API key or cloud provider: pay per token.

Claude plans use a five-hour session limit and a weekly limit. How those work, and how to see what’s left, is in Claude Code usage limits.

Permissions and sandboxing

This is the biggest day-to-day difference.

Codex sandboxes commands at the OS level by default. Locally it starts in workspace-write: it can edit files and run routine commands inside the project, and asks before going further. The other modes are read-only and danger-full-access. Codex uses Seatbelt on macOS, bubblewrap on Linux and WSL2, and a native sandbox on Windows. An “Approve for me” option sends eligible requests to a reviewer agent and only asks you about risky ones (sandboxing docs).

Claude Code gates actions with permission modes. Since v2.1.283, interactive sessions start in auto mode, where a classifier reviews actions instead of you. Manual (default) asks before edits and commands, acceptEdits auto-approves edits, plan only reads, and bypassPermissions is for containers. An OS sandbox (Seatbelt on macOS, bubblewrap on Linux and WSL2) is available for shell commands but off by default: run /sandbox to turn it on (permission modes).

So out of the box, Codex leans on the sandbox and Claude Code leans on the classifier and your answers. Either way, you will still get permission prompts, and an agent that asks while you’re away just waits. Claude Code permission prompts covers the settings that cut them down.

Config files and extensibility

Codex only reads CLAUDE.md if you add it to project_doc_fallback_filenames. Claude Code reads an AGENTS.md when there’s no CLAUDE.md. If you use both agents, keeping one AGENTS.md is the least work.

Hooks in both are the same idea: a script that runs on events like before a tool call, on a permission request, or when the agent stops. See Claude Code hooks for examples.

Where Codex is ahead

• Open source CLI. You can read, patch and audit the agent (Apache-2.0, about 130k GitHub stars).
• Sandbox by default. OS-level isolation is on from the first run on all three desktop systems, including native Windows.
• One subscription. If you already pay for ChatGPT Plus, Codex costs nothing extra, and Free and Go get a taste in the desktop app.
• High ceiling on Pro. The Pro tiers have no five-hour limit at the moment.

Where Claude Code is ahead

• Claude models in their own harness. Opus 5.5 is the default on every paid Claude plan, and Anthropic builds both.
• More places to run the models. Bedrock, Google Cloud, Foundry and gateways, with enterprise controls.
• Plugin ecosystem. Marketplaces bundle skills, subagents, hooks and MCP servers into one install.
• JetBrains. A plugin for IntelliJ, PyCharm, WebStorm and the rest.

Pick Codex if…

• You pay for ChatGPT and don’t want a second subscription.
• You want the agent’s source, or must run it on native Windows with a sandbox.
• You prefer GPT-6.1 Sol’s output on your codebase.

Pick Claude Code if…

• You pay for Claude, or your company buys Claude through AWS, Google Cloud or Azure.
• You want Opus 5.5 or Fable, or the plugin marketplace.
• Your team is on JetBrains.

Using both

Many developers run both: Claude Code on one task, Codex on another, often on two plans so one limit doesn’t stop everything. The tax is attention. Two agents in two terminals means two sets of permission prompts and two “done” moments to notice. Run multiple Claude Code agents covers worktrees and tmux for the DIY setup.

Coral is a Mac app (with an iPhone app) that puts Claude Code, Codex (the CLI or the app), Copilot CLI, OpenCode and Grok sessions in one list in your Mac’s notch, the agent that has waited longest first. You answer a permission prompt with one key, or from your iPhone, and press U to see what’s left of your Claude Code and Codex plans side by side. It’s free; Pro adds things like approval rules and instant delivery from the iPhone. See Features.

Questions

Is Codex better than Claude Code?

Neither wins everywhere. Codex has an open source CLI and a sandbox on by default; Claude Code has Claude’s models, more cloud providers and a plugin marketplace. Try both on the same task in your own repo.

Can I use Codex and Claude Code on the same repo?

Yes. Put shared instructions in AGENTS.md, which both read, and give each agent its own git worktree so they don’t edit the same files.

Is the Codex app free?

The ChatGPT desktop app is free to download, and Free and Go plans get GPT-6 Luna in it as it rolls out. Full Codex, including the CLI and GPT-6.1 Sol, starts with ChatGPT Plus at $20 a month.

Does Claude Code work in Codex’s CLI, or the reverse?

No. Each CLI runs its own company’s models. Tools like OpenCode can use models from several vendors in one interface.

Is the Codex CLI open source?

Yes, under the Apache-2.0 license at github.com/openai/codex. Claude Code’s source is not open.