Claude Code is Anthropic's coding agent and runs only Claude models. Codex is OpenAI's, runs OpenAI's GPT-6 models, and its CLI is open source. Both now work in the terminal, your editor, a desktop app, the web and your phone, and both have plan mode, skills, plugins, hooks, subagents, MCP and memory. What still separates them is whose plan you pay for, how each one fences in commands (the Codex sandbox is on by default, Claude Code's is off) and how each runs on Windows. They also work together: OpenAI ships a Codex plugin for Claude Code, and Codex can import your Claude Code setup.
Checked against Claude Code 2.1.322 and Codex CLI 0.167.1, both released on November 10, 2026, using each vendor's own docs and pricing pages.
| Capability | Claude CodeAnthropic · proprietary | CodexOpenAI · CLI under Apache-2.0 |
|---|---|---|
| What it is | Anthropic's coding agent; the source is closed | OpenAI's coding agent; the CLI is open source (Apache-2.0) |
| Where you use it | Terminal, VS Code and JetBrains, the Claude desktop app (macOS, Windows, Linux in beta), the web and the Claude mobile app | Terminal, VS Code, Cursor and Windsurf (Xcode and JetBrains have their own Codex integrations), the ChatGPT desktop app (macOS, Windows, Linux in preview), the web and the ChatGPT iOS app |
| Models | Claude models only, through Anthropic, Amazon Bedrock, Google Cloud or Microsoft Foundry | OpenAI's GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, depending on your plan; with an API key, the models your API account has |
| Plans that include it | Claude Pro ($20 a month), Max ($100), Ultra ($200), Team or Enterprise, or an API key; the Free plan does not include it | ChatGPT Plus ($20 a month) for the CLI, editor, web and iOS; Pro at $100, $200 or $500; Business or Enterprise; or an API key. OpenAI lists Free and Go for the desktop app with GPT-6 Luna, still rolling out |
| Approvals | Permission modes from Manual to auto, which runs background safety checks and is the starting mode in the terminal and VS Code; deny rules hold in every mode | Ask for approval (recommended), Approve for me (automatic review) or Full access; in a git folder the CLI starts with workspace writes and on-request approvals |
| Sandbox | Off by default; /sandbox turns on an OS sandbox for shell commands on macOS, Linux and WSL2 | On by default for every command it spawns, on macOS, Linux, WSL2 and native Windows; .git and .codex stay read-only |
| Windows | Runs natively; shell commands run unsandboxed unless you use WSL2 | Runs natively in PowerShell with its own Windows sandbox, or in WSL2 |
| Instructions and memory | CLAUDE.md, or a repository's AGENTS.md, plus auto memory notes it writes itself, on by default | AGENTS.md, plus local Codex memories you can switch on or off per machine |
| Extending it | Skills, plugins and marketplaces, subagents, agent teams (experimental), hooks and MCP servers | Skills, plugins that bundle skills, MCP servers and hooks, custom subagents, and the Codex SDK |
| Planning and undo | Plan mode, and checkpoints you can roll back with /rewind | Plan mode with /plan; for rollback, OpenAI suggests git checkpoints before and after a task |
| Working while you are away | Cloud sessions on the web, cloud scheduled tasks (research preview), desktop scheduled tasks and Remote Control from your phone | Codex Cloud tasks that keep running while your computer sleeps, scheduled tasks, and the ChatGPT iOS app to follow and approve work |
| Code review and team tools | Code Review, GitHub Actions, GitLab CI/CD and Slack | Automatic GitHub code review (GitLab in preview), a GitHub Action and @ChatGPT in Slack |
Neither is a complete security boundary: Claude Code's sandbox covers shell commands only, and Full access in Codex removes its sandbox. Many comparisons online are out of date: they cite GPT-5-era Codex models (GPT-5.5 left Codex for ChatGPT sign-in on October 14, 2026), count hooks as a Claude Code advantage although Codex has them too, or treat Claude Code as a terminal tool only.
This comparison was last updated on November 12, 2026; check each vendor's site for current terms.
Many people who search this end up running both. OpenAI supports that directly: it ships a plugin that calls Codex from Claude Code, and Codex can import your Claude Code setup.
OpenAI's Codex plugin for Claude Code adds /codex:review for a read-only review, /codex:adversarial-review to challenge a design decision and /codex:rescue to hand Codex a task. It runs the Codex CLI on your machine, needs a ChatGPT plan or an OpenAI API key, and counts against your Codex usage limits.
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup/codex:transfer hands the current Claude Code conversation to Codex; then run the codex resume command it prints. To bring everything, /import in the Codex CLI turns CLAUDE.md into AGENTS.md and settings.json into config.toml, and copies skills, plugins, MCP servers, hooks, subagents and up to 50 chats from the last 30 days. Your Claude Code setup stays as it was.
# in Claude Code
/codex:transfer
# then, in your terminal
codex resume <session-id>
# or bring your whole setup (in an idle Codex CLI session)
/importCodex reads AGENTS.md before it starts work, and Claude Code can read a repository's AGENTS.md in place of CLAUDE.md, so one file of project rules can serve both. Older guides that add Codex to Claude Code as an MCP server with codex mcp-server no longer work: OpenAI removed that command, and the plugin is the supported route.
Orkas runs Claude Code and Codex as two local agents in the same chat, next to OpenCode, OpenClaw and Hermes. The next table lists what Orkas can and cannot hand each of them.
Step-by-step guide: how to run Claude Code and Codex together →
These rows come from Orkas's adapter code for each CLI, which is the same in Orkas 2.0 and the next release. They describe what a host can do with each agent, not how well either one works.
| In Orkas | Claude Code | Codex |
|---|---|---|
| How Orkas starts it | claude -p, with streaming JSON in and out | codex app-server, over JSON-RPC |
| Project folder | Bound to the project folder you choose | Bound to the project folder you choose |
| The next turn | Continues the same session | Continues the same Codex thread |
| Standing instructions | Sent again with every run | Saved in the Codex thread and sent again only when they change |
| Permission level | Use CLI default, Ask for permission (you approve in Orkas) or Full access | Use CLI default, Ask for permission (you approve in Orkas) or Full access |
| While it runs | Shows progress as it works, and you can send a message mid-run | Shows progress as it works; a message you send mid-run joins the current turn |
Orkas checks for Claude Code 2.0.0 or later and Codex 0.145.0 or later, finds both on your PATH, can update either one, and installs neither. Each agent keeps its own sign-in, plan and usage limits. Orkas skills, the services you connected and each agent's Orkas memory reach both in every run.
Orkas is a free, MIT-licensed desktop app for macOS and Windows. It runs the Claude Code and Codex you already have, plus OpenCode, OpenClaw and Hermes, as agents in one chat.
Install Claude Code and Codex first and sign each one in with its own plan or API key. Orkas finds them on your PATH and needs Codex 0.145.0 or later.
Open Connect agents and add both.
Set approvals in AI Team > Claude Code > Runtime settings > Permission level, and the same for Codex. Then, in the same chat, @-mention one to plan or review and the other to build.
What Orkas does not do here: it does not install either CLI, sign you in or add usage to either plan. Cloud tasks, Remote Control and the ChatGPT apps stay with each vendor.
Neither wins outright, and most head-to-head tests online predate the current models. The two run in the same places and share most features, so decide on three things: which plan you already pay for, whether you want Claude or OpenAI models, and how much sandboxing you need, especially on native Windows. Many developers use both, one to plan or review and the other to build.
Both start at $20 a month: Claude Pro for Claude Code, ChatGPT Plus for Codex. Above that, Claude has Max at $100 and Ultra at $200, and ChatGPT Pro costs $100, $200 or $500. Claude's Free plan does not include Claude Code, while OpenAI lists ChatGPT Free and Go for Codex in the desktop app with GPT-6 Luna, still rolling out. How far each plan goes depends on the model and the task, so compare them on your own work.
Not officially. Claude Code runs Claude models only, and Anthropic does not support routing it to other models through a gateway. Codex runs OpenAI models; a custom provider has to offer a Responses API endpoint, and Chat Completions-only endpoints are not supported. To use both model families, run both agents.
Yes. Type /import in an idle Codex CLI session and choose Claude Code. It turns CLAUDE.md into AGENTS.md and settings.json into config.toml, copies skills, plugins, MCP servers, hooks, subagents and up to 50 chats from the last 30 days, and makes your slash commands into skills. The ChatGPT desktop app can also import from Claude Cowork. Your Claude Code setup is not changed, but check imported permissions and hooks before you rely on them.
Yes. OpenAI's Codex plugin for Claude Code lets you run /codex:review or hand over a task with /codex:rescue without leaving Claude Code. Orkas runs both as local agents in one desktop window, each in the project folder you choose, with its own session, progress and approvals, so you can @-mention one to plan or review and the other to build.
Orkas is free and MIT licensed. It drives Claude Code, Codex, OpenCode, OpenClaw and Hermes as local sessions on your own machine.