Claude Code vs Codex: Which AI Coding Agent to Pick

The choice in Claude Code vs Codex is less about which model writes better code and more about how you want an AI agent to work beside your team. Claude Code is Anthropic’s agentic coding tool: it runs in your terminal, IDE, desktop app or browser, works on your real files, and asks before it acts unless you tell it not to. Codex is OpenAI’s equivalent: a CLI, IDE extension, desktop app and cloud service bundled with ChatGPT plans, built around an OS-level sandbox and hand-off to cloud containers. Both read a repo, edit files, run commands and open pull requests. They differ in pricing, defaults, safety model and how much of the work you delegate versus supervise.

This guide compares the two on the points a founder or ops lead actually has to decide: what each costs per seat and per task, how they behave out of the box, how to install and lock them down, where each one wins, and what to watch out for. Every number links to its source. Where sources disagree, we say so.

Claude Code vs Codex at a glance

If you only read one section, read this one. Both tools are mature coding agents that run the same basic loop: read the codebase, plan, edit files, run tests, iterate, and hand back a diff or pull request. The practical differences sit in four places.

• Default posture. Claude Code starts in a Manual permission mode that prompts on the first use of each tool, so it feels collaborative. Codex starts in an “Auto” preset that combines a workspace-write sandbox with on-request approvals, so commands inside the workspace run without asking and network access is off by default.
• Where the work runs. Both run locally. Both also run in the cloud: Claude Code on the web and in scheduled cloud tasks, Codex in Codex Cloud environments where each task gets its own workspace.
• Price of entry. Both start at $20 a month for an individual paid plan. Codex also appears on ChatGPT Free and an $8 Go plan, per OpenAI’s Codex pricing page. Claude Code is not included on Claude’s free plan, per Anthropic’s pricing page.
• Cost and speed per task. In the independent Artificial Analysis Coding Agent Index, Claude Code with its top model scored higher (62 vs 55), while Codex was cheaper per task ($7.04 vs $11.42) and faster (24.8 minutes vs about an hour).

Short version: pick Claude Code when you want an agent you can shape with hooks, skills, subagents and permission rules, and when task quality on hard problems matters more than cost per task. Pick Codex when your team already pays for ChatGPT, you want strong sandbox defaults, and you plan to delegate well-defined tickets in bulk. Many teams run both, and the rest of this article shows how to decide for your own workload rather than taking anyone’s verdict on trust.

How Claude Code works

Claude Code is an agent harness around Anthropic’s Claude models. Anthropic’s overview documentation lists the surfaces: a terminal CLI, VS Code and JetBrains extensions, a desktop app, the web at claude.ai/code, Slack, and the Claude mobile apps. Each surface connects to the same engine, so the project’s CLAUDE.md, settings and MCP servers carry across. If a repo already has an AGENTS.md file for other agents, Claude Code can read that instead of a CLAUDE.md.

The part that sets Claude Code apart is how much of its behaviour you can program. Hooks run shell commands on lifecycle events, such as formatting every edited file. Skills package repeatable procedures that load only when needed. Subagents take a focused job in their own context window and return a summary. MCP connects it to tools like Jira, Google Drive or your own APIs. We covered how these fit together in Claude Code hooks vs skills vs subagents.

Permissions are rules, not suggestions. The permissions documentation describes six modes: default (Manual), acceptEdits, plan, auto (a background classifier reviews actions instead of you), dontAsk, and bypassPermissions, which Anthropic says to use only in isolated containers or VMs. Allow and deny rules such as Bash(npm run *) live in a checked-in settings.json, so a whole team shares the same limits. For OS-level enforcement there is a separate sandbox setting.

Claude Code also runs headless. claude -p "prompt" executes one task and exits, with --output-format json returning the result plus a cost estimate, per the non-interactive mode docs. That is the basis for CI jobs, scheduled tasks and pipelines like the one that writes this blog.

How OpenAI Codex works

Codex is OpenAI’s coding agent, sold as part of ChatGPT. It runs as a CLI, an IDE extension, inside the ChatGPT desktop and mobile apps, and as Codex Cloud. The CLI is open source under the Apache-2.0 licence in the openai/codex GitHub repository, which matters if your security team wants to read the code that runs on developer laptops.

The safety model has two separate controls, described in OpenAI’s approvals and security guide. The sandbox decides what the agent can technically do: read-only, workspace-write or danger-full-access. The approval policy decides when it must stop and ask: on-request, never, or a granular policy that keeps chosen prompt categories interactive. Locally the sandbox is enforced by the operating system, so even a command the model writes cannot reach files outside the workspace or the network unless you allow it. In a folder without version control, Codex launches read-only.

Codex Cloud is the delegation side. You define an environment once (repositories, dependencies, network allowlist, secrets) and then start tasks from the web, desktop or mobile app. According to the Codex Cloud documentation, each task has its own workspace and keeps working while your computer is asleep; you review the diff and test results, then commit or open a pull request.

Project instructions go in AGENTS.md, which /init generates, per the Codex CLI docs. AGENTS.md is an open format now stewarded by the Agentic AI Foundation under the Linux Foundation and used by over 60,000 open-source projects, according to agents.md. Because Claude Code can read it too, one instruction file can serve both agents.

Feature comparison table

The table compares the two on the features buyers ask about most. Where a cell says “yes” for both, the difference is usually in defaults or depth rather than presence, so read the notes under the table before treating any row as a tiebreaker.

Two notes. First, the Codex documentation index lists hooks, subagents and skills too, so the old “Claude Code has more features” argument is weaker than it was a year ago; the gap now is in how mature and documented each piece is. Second, the safety rows reflect defaults, not ceilings. You can make Claude Code as locked down as Codex with a sandbox and deny rules, and you can make Codex as permissive as bypassPermissions with --yolo. The defaults matter because most people never change them.

Pricing, usage limits and a worked cost example

Both vendors bundle the agent into a chat subscription and meter it with rolling windows, so “unlimited” is never on offer. On the Claude side, Anthropic’s pricing lists Pro at $20 per month billed monthly ($17 billed annually), Max from $100 per month at 5x or 20x Pro usage per five-hour session, Team Standard seats at $25 monthly and Premium seats at $125 monthly, and Enterprise at $20 per seat plus usage at API rates with a 20-seat minimum. Free does not include Claude Code.

On the OpenAI side, the Codex pricing page lists Free and Go ($8) with a small model, Plus at $20 with published ranges per five hours (for example 15 to 160 local messages on GPT-6.1 Sol), Pro from $100 with tiers at $100, $200 and $500 and no five-hour limit, and Business at $25 per user per month billed monthly. Plus and Business users can buy credits to extend usage. OpenAI also lets you run Codex on an API key at pay-per-use rates.

Now a worked example for a five-developer agency. On subscriptions alone, five Codex Business seats cost 5 × $25 = $125 a month, and five Claude Team Premium seats cost 5 × $125 = $625 a month, using the list prices above. That gap is real, but it is not the whole story, because heavy users hit limits and buy more. Anthropic’s cost documentation says enterprise deployments average about $13 per developer per active day and $150 to $250 per developer per month, so a budget of $750 to $1,250 a month for five heavy Claude Code users is a realistic ceiling to plan against.

Per-task numbers tell a similar story. Using the Artificial Analysis averages of $11.42 per task for Claude Code and $7.04 for Codex on their hardest settings, 100 delegated tasks a month would cost about $1,142 and $704 respectively at API rates. Codex also used far fewer tokens per task in that test (3.2 million vs 14.5 million). If your work is a steady stream of well-specified tickets, that difference compounds. If a single wrong fix costs you a client, the higher success rate may be worth more than the $4 saved per task.

Performance: what the benchmarks do and do not show

Benchmarks for coding agents test the whole harness, not just the model, and they change every few weeks. The most useful public comparison we found is Artificial Analysis’s Claude Code vs Codex page, which runs both agents on the same tasks. In that index, Claude Code with Opus 5.5 scored 62 and Codex with GPT-6 Astra scored 55. Claude Code also led on its Terminal-Bench 4.0 component (63% vs 56%). Codex finished tasks in 24.8 minutes on average against about an hour for Claude Code, using 39 turns per task instead of 146.

How to read that: Claude Code explores more, checks more and spends more tokens doing it, which tends to pay off on ambiguous or multi-step work. Codex is more direct, which pays off on tasks with a clear definition of done. That matches the framing in DataCamp’s comparison, which describes Codex as built for delegation and Claude Code for collaboration.

What the benchmarks do not show is how either tool performs on your repository, with your test suite, your conventions and your messy legacy module. They also mix model and harness: Claude Code can run several Claude models and Codex several OpenAI models, and the rank order shifts with the model and effort level chosen. Treat any public score as a reason to run a trial, not as the answer. The section on testing both below shows a cheap way to do that in a week.

Step-by-step: install and lock down both tools

Installing both takes under ten minutes. The commands below come from the official docs; run them on a test repository first, not on production code.

1. Install Claude Code

On macOS, Linux or WSL, Anthropic’s installation docs recommend the native installer, which updates itself:

curl -fsSL https://claude.ai/install.sh | bash
cd your-project
claude

Homebrew users can run brew install --cask claude-code instead. On first launch you log in with a Claude subscription or a Console API key.

2. Install Codex

The Codex repository lists an install script, npm and Homebrew:

curl -fsSL https://chatgpt.com/codex/install.sh | sh
# or
npm install -g @openai/codex
cd your-project
codex

Choose “Sign in with ChatGPT” to use your plan’s included usage, or an API key to pay per token.

3. Share one instruction file

Write your build, test and style rules once in AGENTS.md. Codex reads it natively, and Claude Code reads it when there is no CLAUDE.md. Keep it short: commands to run tests, folders not to touch, and how to name branches.

4. Set safe defaults for Claude Code

Check a .claude/settings.json into the repo so every developer gets the same rules. This example, adapted from the permissions docs, lets the agent run npm scripts and commit, but refuses pushes:

{
  "permissions": {
    "allow": ["Bash(npm run *)", "Bash(git commit *)"],
    "deny": ["Bash(git push *)"]
  }
}

5. Set safe defaults for Codex

In ~/.codex/config.toml, the approvals guide uses these keys. Leave network access off unless a task needs it:

sandbox_mode = "workspace-write"
approval_policy = "on-request"

[sandbox_workspace_write]
network_access = false

6. Run both headless in CI

For a nightly job or pull request check, use the non-interactive modes. Note that codex exec runs in a read-only sandbox by default and --full-auto is no longer supported, per the non-interactive docs:

claude -p "Run the test suite and fix any failures" --allowedTools "Bash,Read,Edit" --output-format json
codex exec --sandbox workspace-write --json "Run the test suite and fix any failures"

Who should pick which

The right choice depends more on your team and your existing subscriptions than on benchmark scores. Here is how we would decide for the kinds of businesses that read this blog.

Pick Claude Code if

• You want to build repeatable automations around the agent, not only use it for coding. Hooks, skills, subagents and the Agent SDK make it a general agent platform; this blog’s daily writer is a Claude Code scheduled task, and our overview of Anthropic Claude for automation covers the wider picture.
• Your work is ambiguous: refactors across many files, migrating a legacy codebase, or tasks where planning first saves rework.
• You want a human in the loop by default. Manual mode makes approvals the norm, which suits teams new to agents. See our guide to human-in-the-loop agentic AI for how to design those checkpoints.

Pick Codex if

• Your company already pays for ChatGPT Business or Enterprise. The agent is included, so the marginal cost of a trial is zero.
• You want to hand off a backlog of well-defined tickets and review the pull requests later. Codex Cloud is built for that pattern.
• Your security team prefers an open-source client and a sandbox that is on by default.

Not a fit for either

If nobody on the team can read a diff, neither tool is safe to run unsupervised on production code. Start with a no-code automation platform, or bring in someone who can review agent output. Our piece on coding with LLMs in 2026 covers the review skills a team needs first.

Common mistakes and how to avoid them

Most failed rollouts of coding agents trace back to the same handful of errors. None of them is specific to one vendor, but each tool makes some of them easier to fall into.

1. Turning off every guard on day one. bypassPermissions in Claude Code and --yolo in Codex remove all prompts. Both vendors say to use them only in isolated environments. Why it matters: an agent that can run any command can also delete a branch or leak a secret. Use auto or workspace-write modes instead.
2. Judging on one demo task. A single to-do app tells you little. Run the same ten real tickets through both tools before deciding, as described in the next section.
3. Ignoring usage windows. Both plans reset on rolling five-hour windows and may add weekly caps. Teams that budget only the seat price get surprised mid-sprint. Check /usage in Claude Code and /status in Codex during the trial.
4. Letting sessions grow forever. Long sessions resend the whole conversation on every request. Anthropic’s cost guide recommends /clear between unrelated tasks and choosing a cheaper model for routine work; the same habit cuts Codex usage too.
5. Maintaining two instruction files. A CLAUDE.md and an AGENTS.md that drift apart produce two agents with different rules. Keep one source of truth.
6. Skipping review. Treat agent pull requests like a new hire’s: tests must pass and a person reads the diff. Speed gains vanish the first time an unreviewed change breaks billing.

A real example: how we would run a one-week trial

Benchmarks give you a prior; your own repo gives you the answer. Here is the trial we run with clients who cannot decide, and it costs less than one day of developer time.

1. Pick ten closed tickets from the last quarter that had a clear definition of done: three bug fixes, three small features, two refactors, two test-writing tasks. Because they are closed, you already know the correct answer.
2. Branch from the commit before each fix and give both agents the original ticket text, nothing more.
3. Run each tool in its default mode with the shared AGENTS.md. Record wall-clock time, how many times it asked for approval, whether tests passed, and how much of the diff a reviewer would keep.
4. Record cost. Claude Code’s JSON output includes an estimated total_cost_usd, per the headless docs; on a ChatGPT plan, note how much of the five-hour window each task used.
5. Score and decide. A tool that solves eight of ten with clean diffs at higher cost usually beats one that solves six cheaply, because every failure costs a developer an hour.

In trials like this, the result often splits by task type: one tool wins refactors, the other wins small, crisp fixes. That is why many teams end up running both, with an agreed rule for which tickets go where. If you are coordinating several agents, our guide to orchestrating AI agents for business covers routing and hand-offs.

Alternatives worth a look

Claude Code and Codex are not the only options, and for some teams neither is the best first step. The main alternatives fall into three groups.

• AI-first editors such as Cursor and Google Antigravity put the agent inside the IDE with a visual diff, which suits developers who live in an editor rather than a terminal. We compared them in Antigravity vs Cursor vs Claude Code.
• Other terminal agents such as Gemini CLI, Aider, goose and opencode. Several of them read AGENTS.md too, according to the AGENTS.md project, so switching costs stay low.
• No-code automation platforms such as n8n, Make and Zapier. If the goal is automating a business process rather than writing software, a workflow tool with an AI step is usually cheaper and easier to hand over to an ops team.

When to use an alternative instead: choose an editor-based agent when your developers want inline suggestions as they type; choose a no-code platform when the “code” is really a chain of API calls between SaaS tools. Choose Claude Code or Codex when the job is changing a real codebase, end to end, with tests.

Decision checklist

Use this list in your buying meeting. If you can answer each line, you have enough to pick one, both, or neither with confidence.

• Which chat subscription does the company already pay for, and does it include the agent?
• Which tasks will the agent do: crisp tickets, ambiguous refactors, or non-coding automation?
• Do we want approvals by default (Claude Code Manual mode) or a sandbox by default (Codex Auto)?
• Who reviews agent pull requests, and how fast?
• What is the monthly budget per developer, including overage beyond the seat price?
• Is network access needed during tasks, and which domains?
• Do we have one shared AGENTS.md with test commands and no-go folders?
• Have we run the ten-ticket trial and recorded time, cost and accepted diff share?
• Who owns permission settings in version control?
• When do we re-evaluate? Pricing and models change every quarter.

FAQ

Can I use Claude Code and Codex in the same repository?

Yes. They do not conflict, and both can read the same AGENTS.md. Run them on separate branches or git worktrees so their edits do not collide, and keep one instruction file so their rules match.

Is Codex free?

Codex appears on ChatGPT Free and Go with a smaller model, subject to rollout, according to OpenAI’s pricing page. For serious daily use you need Plus, Business or Pro, or an API key billed per token.

Which is better for non-developers?

Neither is designed for people who cannot read code. If you must choose, the desktop apps of both are friendlier than the terminal. For business automations without code, a workflow platform is a better fit.

Do these tools send my code to the vendor?

Yes. Both send the context they need to the vendor’s models to work. Business and Enterprise plans add admin controls and data terms; check them with your vendor before using either on client code.

Which is better on Reddit?

Opinions split along the same lines as the benchmarks: Claude Code gets credit for depth and features, Codex for value and speed. Most threads end with “use both”. Treat forum verdicts as anecdotes and run your own trial.