I build Morphy, an app that runs both of these side by side, so I don't need you to pick one. That makes this easier to write honestly: both are excellent, they're better at different things, and in 2026 the two companies made it officially easy to use them together.
What are they?
Codex is OpenAI's coding agent. It runs in your terminal (the Codex CLI, which is open source), in VS Code and its forks, JetBrains IDEs and Xcode, inside the ChatGPT desktop app since July 2026, and in the cloud, where tasks keep running without your computer. It uses OpenAI's GPT-6 models.
Claude Code is Anthropic's coding agent. It runs in your terminal, in VS Code, Cursor and JetBrains IDEs, in the Code tab of the Claude desktop app, on the web and from your phone. It uses Anthropic's Claude models.
Both read your project, run commands, edit files, use MCP servers and skills, and can split work across subagents. The differences are in the details.
How do they compare?
| Codex | Claude Code | |
|---|---|---|
| Made by | OpenAI | Anthropic |
| Models | GPT-6 Astra (top), GPT-6.1 Sol (recommended, cheaper), GPT-6 Sol and Luna | Opus 5.5 (default), Sonnet 5.5, Haiku 5.5, Fable 5.1 |
| Cheapest plan that includes it | ChatGPT Free, in the desktop app (with GPT-6 Luna, as it rolls out) | Claude Pro, $17 a month billed yearly or $20 monthly |
| Main paid plans | Plus $20; Pro $100, $200 or $500 a month | Pro $17–20; Max $100 or $200 a month |
| Usage limits | Plus: 5-hour windows plus weekly limits; Pro: no 5-hour limit | Rolling 5-hour windows plus weekly limits, shared with Claude chat |
| Runs on | Terminal, IDEs, ChatGPT desktop app (Windows, Mac, Linux preview), cloud, phone | Terminal, IDEs, Claude desktop app (Windows, Mac, Linux beta), web, phone |
| Safety by default | A sandbox: no network, writes only in your project; native on Windows | Auto mode: a classifier reviews each action; the command sandbox is opt-in |
| Project instructions | AGENTS.md | CLAUDE.md, and AGENTS.md when there's no CLAUDE.md |
| Pull request reviews | Managed GitHub reviews from Plus, nothing to set up | A GitHub Action review workflow on Pro and up; managed reviews on Team and Enterprise (preview); /code-review locally |
| Open source | The CLI, Apache-2.0 | No (its sandbox runtime is Apache-2.0; the GitHub Action and Agent SDK for Python are MIT) |
| Trains on your work (personal plans) | Yes, unless you turn it off | Only if you allow it in Claude's privacy settings |
Where is Codex better?
- It's cheaper to start. ChatGPT's free and Go ($8) plans include Codex in the desktop app, with the smaller GPT-6 Luna model as it rolls out. Claude's free plan doesn't include Claude Code at all.
- Sandboxed by default. Out of the box, Codex can't reach the network and can only write inside your project, and it asks before stepping outside. That sandbox works natively on Windows. Claude Code's own command sandbox is off by default, and on native Windows its commands run unsandboxed.
- Pull request reviews with nothing to set up. From ChatGPT Plus, mention
@codex reviewon GitHub or turn on automatic reviews. Claude Pro can do automatic reviews too, through the Claude Code GitHub Action, but that's a workflow you add to your repository and it runs on your GitHub Actions minutes; Anthropic's managed review is for Team and Enterprise. - More room at the top. ChatGPT Pro has no 5-hour limit (weekly limits may still apply), and its $500 tier adds an Ultrafast mode for GPT-6 Astra and GPT-6.1 Sol. Ultrafast uses eight times the usage, Fast two and a half.
- Cheaper to run. On Terminal-Bench 4.0, Vals.ai measured GPT-6.1 Sol at $1.72 a task against $13.20 for Opus 5.5. Per token, GPT-6.1 Sol costs the same as Claude Sonnet 5.5 ($2 per million input tokens, $10 output), and OpenAI's top model, GPT-6 Astra, costs more than Opus ($10 and $50 against $4 and $20).
- Open source. You can read, fork and build the Codex CLI yourself.
Where is Claude Code better?
- Stronger on the benchmark that matters most right now. On Terminal-Bench 4.0, which measures agents doing real work in a terminal, Claude Code with Opus 5.5 scores 64.85 on the official leaderboard, against 58.18 for Codex with GPT-6 Astra or GPT-6.1 Sol, all at maximum effort. Vals.ai, running every model in the same harness, puts Opus 5.5 at 65.15 and GPT-6 Astra at 59.60. That lead costs more, though: see the price per task under Codex.
- Deeper customization. Claude Code has nearly three times as many hook events as Codex (33 against 12) and experimental agent teams, on top of the subagents, skills, plugins, MCP and scheduled cloud tasks that both have (Claude's routines are still a research preview).
- Fewer prompts out of the box. In the terminal and VS Code, auto mode is now the default: a classifier approves routine actions, so long tasks run without approving every step. (Codex gets close with its sandbox plus Approve for me, and Anthropic itself recommends adding a sandbox for truly hands-off work.)
- It's in the Claude app you may already use, next to your chats, with Projects (a beta on Pro and Max, still rolling out). Pick Opus for hard work, Sonnet for most of it and Haiku for quick jobs; Fable 5.1 costs extra usage on Pro and is capped on Max.
What do they cost, and how far does a plan go?
Both are included in the chat subscription, and both count usage in windows rather than per message.
- ChatGPT (for Codex): Free and Go include it in the desktop app. Plus ($20) adds the CLI, cloud tasks and GitHub reviews, with 5-hour windows and weekly limits. Pro ($100, $200 or $500) has no 5-hour limit. Pro $200's allowance is dropping: for new subscriptions now, and for existing ones after October 29.
- Claude (for Claude Code): Pro ($17–20) includes it. Max ($100 or $200) gives five or twenty times Pro's usage, plus $100 or $200 a month in API credits. Claude Code and Claude chat draw from the same pool.
Running several agents at once uses a plan up several times as fast, with either one. That's where having both helps: when one plan runs out, the next agent can run on the other.
A word on benchmarks
Benchmarks change monthly and measure a model inside one particular harness, not your codebase. The older SWE-bench leaderboards haven't caught up with these models at all, so Terminal-Bench 4.0 is the best independent comparison right now, and it favours Claude by about six points. Try both on your own work before you commit to a $100 plan.
What about privacy?
On personal plans, OpenAI may train on your Codex work unless you switch it off; Anthropic only does if you allow it:
- OpenAI: turn off Improve the model for everyone in ChatGPT's data controls, and check Include environments in Codex's own settings, which is separate. Business, Enterprise, Edu and API use isn't used for training by default.
- Anthropic: it's your choice under Help improve our AI models in Claude's privacy settings, and it covers Claude Code too. With it on, data is kept for five years; with it off, 30 days. Team and Enterprise aren't used for training.
Can you use both?
Yes, and it's now officially supported in both directions:
- One instruction file. Keep your project's instructions in
AGENTS.md: Codex reads it, and Claude Code reads it too when there's noCLAUDE.md(a setting makes it read both). - Import each other's setup. Codex's
/importbrings over your Claude Code setup (instructions, settings, skills, hooks, subagents, MCP servers), and Claude Code's/import codexbrings Codex's instruction files, MCP servers, commands, subagents and skills the other way. - One builds, the other reviews. OpenAI's own Codex plugin for Claude Code adds
/codex:reviewand/codex:adversarial-reviewinside Claude Code, so Codex checks what Claude did, and/codex:rescue, which hands a task to Codex.
Or run them side by side. In Morphy, a Claude chat and a Codex chat can work on the same project next to each other on one canvas: pick the AI from each chat's title bar, compare what they come back with, and watch both plans' usage in the corner. Morphy's free plan runs three agents at a time; see Connect your AI to sign in to both.
Which should you pick?
- No budget yet: Codex, on ChatGPT's free plan.
- On Windows, without WSL: Codex, for its native sandbox.
- You want pull request reviews on a $20 plan with nothing to set up: Codex, on ChatGPT Plus.
- You want the best results on long agent tasks today: Claude Code with Opus 5.5.
- You want to shape the agent with hooks, subagents and routines: Claude Code.
- You already pay for one of them: start there; the other is one import away.
- You keep hitting limits: both, with one
AGENTS.md, and run them side by side.
For every other tool that runs agents like these in parallel, see the full roundup.
References
- OpenAI: Codex pricing and limits, models, speed modes, hooks, CLI, IDE extension, Codex CLI on GitHub, Codex plugin for Claude Code, how your data is used
- Anthropic: Claude Code setup, GitHub Actions, permission modes, CLAUDE.md and AGENTS.md, model configuration, data usage, Claude pricing
- Benchmarks: Terminal-Bench leaderboard, Vals.ai Terminal-Bench 4.0
