Claude Code or Codex: which coding agent to choose in 2026?
Comparison published on October 10, 2026, established from the official documentation of Anthropic and OpenAI consulted that day, and the interfaces of Claude Code and Codex opened on the same day in our accounts (Claude Max subscription, professional ChatGPT workspace).
Claude Code or Codex? These are today the two main code agents: a task is described to them in plain language, they read the project, modify the files, launch the commands and tests, and then propose the result. Both work in the terminal, code editor, desktop application, browser, and mobile. Claude Code stands out for its deep customization options and automatic authorization mode; Codex for its integration with the ChatGPT subscription, its strict default segmentation, and its speed on the more expensive plans. The choice depends primarily on your existing subscription and the level of autonomy you want to grant the agent.
The verdict in five lines
- You already have a ChatGPT subscription: Codex is included in Plus, and is available in a limited version for free.
- You want to fine-tune the agent to your team: Claude Code, with its hooks, MCP connectors, skills and agent development kit.
- Default security: Codex works in a sandbox without network access; Claude Code relies on a classifier that blocks risky actions without interrupting you.
- Background and planned tasks: They both know how to work in the cloud, in parallel and at fixed times.
- Speed: Codex offers an Ultrafast mode up to 8 times faster, reserved for high-end offers; Claude Code offers a fast mode up to 2,5 times faster on Opus 5.5.
The comparative table
| Criterion | Claude Code (Anthropic) | Codex (OpenAI) |
|---|---|---|
| Where to use it | Terminal, VS Code, JetBrains, desktop app (macOS, Windows, Linux beta), browser (claude.ai/code), mobile, Slack, Chrome | Terminal, editor extension, desktop application (macOS, Windows), cloud in ChatGPT, mobile, GitHub, Slack |
| Models | Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1 (1 million context tokens), with effort adjustment | GPT-6.1 Sol, GPT-6 Sol, GPT-6 Astra, GPT-6 Luna, with effort adjustment |
| Project instruction file | CLAUDE.md, and reading of an existing AGENTS.md; automatic memory management between sessions | AGENTS.md |
| Permissions | Auto mode (a classifier blocks risky actions), Manual mode, Change acceptance mode, Plan mode | Read-only sandbox, write-in-workspace, or full access; approval policies; automatic agent review |
| Default network access | According to the authorizations granted | Off by default, must be explicitly enabled |
| Go back | Automatic checkpoints before each file modification | Review of changes by discussion thread, isolated worktrees |
| Extensions | MCP, skills, hooks, plugins, sub-agents, Agent SDK | Plugins (GitHub, Slack, Gmail, Google Drive, Notion, Figma…), skills |
| parallel work | Sub-agents, background agents, worktrees, workflows orchestrating dozens of sub-agents | Parallel agents grouped by project, worktrees integrated into the application |
| Scheduled Tasks | Routines in the cloud (schedule, GitHub event, API call) and tasks scheduled on the computer | Automations within the application, results in a review queue |
| Code review | Automatic review of pull requests on GitHub, GitHub Actions, and GitLab CI/CD | Automatic review of pull requests on GitHub |
| Quick mode | Fast mode on Opus 5.5, up to 2,5 times faster | Ultrafast, up to 8x faster, on Pro at $500 and select Enterprise offerings |
| Free offer | No: paid subscription required | Yes, limited (GPT-6 Luna in the desktop application) |
| Individual offer | Pro: $20 per month ($17 annually); Max: $100 or $200 | Plus: $20 per month; Pro: $100, $200, or $500 |
| Team offer | Team: $25 per seat per month ($20 annually), premium seat $125 ($100) | Business: $25 per seat per month ($20 annually) |
| Other access points | API key, Amazon Bedrock, Google Cloud, Microsoft Foundry | API key (without cloud functions) |
The prices shown are those in dollars displayed by the publishers and may vary in euros depending on VAT. In both cases, the use of the code agent is counted towards the same limits as the chatbot.
What we saw in both interfaces
- Claude Code in the browser (claude.ai/code): We choose a repository and an environment, then we describe the task. The selector offers the model (Opus 5.5 by default on our account), the effort level (medium by default), the authorization mode (Auto) and the fast mode, disabled by default.
- Codex in ChatGPT You choose where to work (in the cloud or remotely on your own machine) and an environment, then you describe the task. The selector offers GPT-6.1 Sol with a medium effort by default, and the "Ask for approval" mode, which requests your agreement before sensitive actions.
The two interfaces are very similar: an input field, a repository, a template, a level of autonomy. The difference lies less in the screen than in the default settings and what can be connected to it.
Two philosophies of security
Codex starts from a closed framework. The agent operates in a sandbox where you choose the level of access: read-only, write access limited to the workspace (the "Auto" preset), or full access, which the documentation advises against. Network access is disabled by default and must be explicitly enabled. Approval requests may be reviewed first by a review agent.
Claude Code starts with fluidity. Its Auto mode, now the default for interactive sessions, delegates authorization requests to a classifier: safe actions execute without interruption, while risky ones are blocked. The Plan mode allows a plan to be proposed without modifying the files, and each file change is preceded by a checkpoint that allows for reverting. Actions on remote systems (databases, deployments) cannot be undone in this way: they are governed by authorization rules.
In a company where the IT department wants to maintain control, Codex's default partitioning provides reassurance. For a developer who wants to delegate lengthy tasks without validating each command, Claude Code's Auto mode saves time.
Customize the agent
Both agents read an instruction file located at the project's root: AGENTS.md for Codex, and CLAUDE.md for Claude Code, which can also read an existing AGENTS.md file. A team can therefore share the same instructions between the two tools.
Beyond that, Claude Code offers more points of interest: hooks that run a command before or after an action (formatting the code after each modification, running tests before a commit), MCP connectors to Jira, Google Drive, or your internal tools, skills shareable, sub-agents specialized and a Agent SDK to build your own agents. Codex focuses on plugins ready to use (GitHub, Slack, Gmail, Google Drive, Notion, Figma, Supabase) and on skills manageable from its application.
Working in parallel and in the background
The two developers have made parallel work their priority for 2026. The Codex application presents itself as a "command center" for agents: each task runs in its own thread, on an isolated copy of the repository (worktree), and scheduled automations submit their results to a review queue. Claude Code offers a view of all ongoing sessions, background agents, workflows capable of orchestrating dozens of sub-agents, and cloud-based routines that are triggered by a schedule, a GitHub event, or an API call, even when the computer is turned off.
The models and the cost
On the subscription side, the two plans are identical: $20 per month for both Claude Pro and ChatGPT Plus, and $25 per seat per month ($20 annually) for both Claude Team and ChatGPT Business. Codex maintains an entry advantage: a limited version is available for free, while Claude Code requires a paid subscription.
In terms of modeling, OpenAI presents GPT-6.1 Sol, released in late September 2026, as being close to its most powerful model (GPT-6 Astra) at a fifth of the price. Anthropic presents Opus 5.5, released on September 22, 2026, as its reference model for agentic code, 40% less expensive than Opus 5 on typical tasks. These figures come from the vendors themselves: you should verify them on your own tasks.
Which one to choose depending on your situation
| Your situation | Our recommendation |
|---|---|
| Team already equipped with ChatGPT Business or Enterprise | Codex, already included |
| Team already equipped with Claude Team or Enterprise | Claude Code, already included |
| IT department demanding on compartmentalization | Codex, closed by default |
| Advanced automation (hooks, CI, custom agents) | Claude Code |
| Model hosting on AWS, Google Cloud or Microsoft | Claude Code, available on these three platforms |
| Discovery without a budget | Codex, with its limited free access |
| Non-developer profiles who want to create tools | Both, via their web interface or desktop application, rather than the terminal |
For the conversational aspect of these two ecosystems, see our comparison. Claude or ChatGPTand to choose the right model depending on the task, The right AI model for each task.
F.A.Q
Claude Code or Codex: which is better for coding?
Neither excels in every area. Both developers publish test results that place them at the top in certain areas. In practice, Claude Code is more customizable and smoother to use independently, while Codex is more compartmentalized by default and included in the ChatGPT subscription.
Claude Code vs Codex: which is cheaper?
Subscriptions are the same price: $20 per month for individuals, $25 per seat per month for teams ($20 annually). Codex also offers limited free access. The actual cost depends primarily on usage, which is calculated based on limits that are renewed every five hours and weekly.
Do you need to know how to program to use them?
Not for starting out: their web versions and desktop applications allow you to describe a task in French and review the result without opening a terminal. But reviewing and validating what the agent produces requires a foundation, especially before a production deployment.
Can both be used on the same project?
Yes. Claude Code can read the AGENTS.md file used by Codex, which allows for the sharing of the same instructions. Some teams assign code review to one person and writing to the other.
Is my code being used to train the models?
On the team and enterprise plans from both publishers (Claude Team and Enterprise, ChatGPT Business and Enterprise), training is not enabled by default. On individual plans, training can be disabled in the settings.
Sources
- Anthropic, Claude Code's documentation: presentation et operation.
- Anthropic, new releases from Claude Code.
- Anthropic, Claude's offers, accessed on October 10, 2026.
- Anthropic, Introducing Claude Opus 5.5, September 22, 2026.
- Anthropic, Use Claude Code with a Pro or Max plan.
- OpenAI, Codex presentation.
- OpenAI, Codex offers and limitations, accessed on October 10, 2026.
- OpenAI, Codex approvals and security.
- OpenAI, Introducing the Codex app.
- OpenAI, Introducing GPT-6.1 Sol.