Claude Code vs Codex: Which to Pick

Both are terminal-first coding agents from the two big model vendors. Codex CLI is Apache-2.0 and comes with ChatGPT plans, starting free. Claude Code is proprietary and comes with Claude plans from $17 a month. The deciding factor is usually which subscription you already pay for, then which vendor's models you trust on your codebase. Facts below were checked on 31 October 2026.

The short version

What each one is

Codex

The README: "Codex CLI is a coding agent from OpenAI that runs locally on your computer." The same README points to three other ways in. If you want Codex in your editor (VS Code, Cursor, Windsurf), install it in your IDE. If you want the desktop experience, run codex app. If you want the cloud-based agent, go to Codex Web at chatgpt.com/codex.

So "Codex" means a family: CLI, IDE extension, desktop app, web and iOS. The pricing page says Plus includes "Codex on the web, in the CLI, in the IDE extension, and on iOS."

The CLI docs list features such as resuming sessions, adding images, delegating focused work to specialized agents, live web search, MCP servers, approval modes with a visible sandbox, and a dedicated review command that reports prioritized findings without modifying your working tree.

Claude Code

Anthropic's docs: an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. It runs in the terminal, in VS Code and JetBrains, in a desktop app and on the web. One engine backs all of them, so a repo's CLAUDE.md, settings and MCP servers work everywhere.

Side-by-side facts

Plans and prices

Codex

OpenAI's pricing page says "ChatGPT Work and Codex are included in your ChatGPT Free, Go, Plus, Pro, Business, Edu, or Enterprise plan," and that they share usage.

Usage is described in local messages per five-hour period for Plus and Standard Business. The page's table gives ranges that vary by model: for example, about 15 to 160 messages for GPT-6.1 Sol, 35 to 285 for GPT-6.1 Terra, and 350 to 3,000 for GPT-6 Luna. OpenAI calls these estimates, not fixed limits, and says cloud tasks may use more of your allowance. Pro plans currently have no five-hour limit. Weekly limits may also apply.

When you hit the limit, Plus and Pro users can buy additional credits. Credits are quoted per million tokens, and the rates differ by model.

Claude Code

Anthropic's cost page gives reference numbers for API usage: an average of about $13 per developer per active day, or $150 to $250 per developer per month, with 90 percent of users under $30 per active day.

Comparing the two

The price ladders are close, with one notable gap at the bottom. At $20 both sell a mid-level plan. OpenAI also has a $0 plan and an $8 Go plan with Codex included, while Claude Code needs at least Pro. At the top, both have $100 and $200 tiers, and OpenAI adds a $500 option.

The meters are alike in kind. Both use rolling windows with weekly caps and estimates that depend on model and task, and both let you pay for more. Neither vendor publishes a fixed "tasks per month" number, so a price comparison without your own usage data is a guess. We did not measure usage on either.

Models

Codex uses OpenAI models. The pricing page lists a series of GPT models, with GPT-6.1 Sol and GPT-6 Luna on Plus. It also notes GPT-5.5 was retired from Codex with ChatGPT sign-in on October 14, 2026, while the OpenAI API and API-key sessions were unaffected. That is a reminder that the model list moves.

Claude Code uses Claude models, with aliases such as opus, sonnet, haiku, fable and best, and an opusplan mode that uses Opus to plan and Sonnet to execute. Third-party providers are supported in the terminal CLI, VS Code and JetBrains.

Neither tool is built to run the other vendor's models as a first-class path. If you want to switch freely, tools like OpenCode are built for that.

Workflow and automation

Scripts and CI

Codex: codex exec runs Codex non-interactively, and OpenAI's docs describe it for scripts and CI. The API-key plan is described as great for automation in shared environments.

Claude Code: claude -p runs a prompt and exits. Anthropic shows it in pipes (git diff main --name-only | claude -p "review these changed files for security issues"), and it has GitHub Actions and GitLab CI/CD integrations for review and issue triage.

Parallel work and delegation

Codex lets you ask it to delegate focused work to specialized agents and bring their findings back into the main session. Claude Code subagents run in their own context windows with a custom system prompt, tool access and permissions, and return a summary. A lead agent can coordinate several, and background agents let you watch multiple sessions.

Review

Codex has a review mode against uncommitted changes, a commit or a base branch that reports findings without touching your working tree. It also has Codex Security for scanning code and reviewing fixes. Claude Code has bundled skills such as /code-review, GitHub Code Review for PRs and CI integrations.

Scheduling and remote

Claude Code's cloud scheduled tasks keep running when your computer is off, and Remote Control, claude --teleport and /desktop move sessions between surfaces. Codex has cloud tasks you can submit to a configured environment and apply to your local repository from the terminal, and the pricing page notes scheduled automations in its agent documentation.

Permissions and safety

Codex's CLI docs describe choosing when it may edit files or run commands without asking, and show the active sandbox and writable roots. Claude Code asks before edits and shell commands except a built-in read-only set, offers plan mode, matches bash rules against whole commands including operators, and has a sandbox with managed settings for organizations.

Both are sandbox-aware. Neither is a substitute for running an agent on a branch with review before merge.

Data handling

Anthropic's data-usage page: for Free, Pro, Max and Ultra accounts there is a setting that lets data be used to improve future models, including Claude Code use, with five-year retention for those who allow it. Team, Enterprise, API and third-party platform use falls under commercial terms with no training on Claude Code inputs unless the customer opts into a program.

For Codex, data terms depend on the ChatGPT plan or API account.

Where Codex is stronger

• An open-source CLI under Apache-2.0.
• Free and $8 entry plans that include Codex.
• A sign-in that reuses an existing ChatGPT subscription, with no extra product to buy.
• A broad set of surfaces, including iOS, plus the editor install for VS Code, Cursor and Windsurf.
• A $500 top tier for the heaviest users.

Where Claude Code is stronger

• The longest-documented hook, skill and subagent system.
• Cloud scheduled tasks that run without your computer.
• One engine across terminal, VS Code, JetBrains, desktop and web, with handoff.
• Usage attribution by skill, subagent, plugin and MCP server in /usage.
• Bedrock and Vertex AI support for teams that buy models through a cloud.

Which to pick

Choose Codex when:

• You already hold a ChatGPT plan.
• Open source matters.
• You want to start at no cost.

Choose Claude Code when:

• You already hold a Claude plan.
• You want hooks and subagents as part of your daily process.
• You schedule agent work.

Use both when:

• You want a second opinion on a hard change. A shared AGENTS.md keeps your instructions in one place, since Codex documents it and Claude Code can read it.

Switching

Move instructions first. Keep an AGENTS.md for Codex and let Claude Code read it, or copy it to CLAUDE.md. Re-add MCP servers using each tool's command. Recreate repeatable workflows as skills, since both document skills.

A short checklist

• Which subscription do you already have? Start there.
• Do you need the code open? Codex.
• Do you need scheduled or CI runs? Both support CI; check scheduling.
• Do you need an iOS app or an editor install? Codex has both.
• Who owns the data setting on individual accounts?

Limits and gotchas

• Both meters are estimates. OpenAI calls its message counts estimates, and says cloud tasks may use more of your allowance. Anthropic's plans use five-hour and weekly windows. Plan for a day you hit the wall.
• Model lists move. OpenAI retired GPT-5.5 from Codex with ChatGPT sign-in on 14 October 2026. A workflow pinned to one model name can break on a vendor's schedule, so pin as little as you can.
• Codex usage is shared. ChatGPT and Codex draw on one allowance, so a long chat session in the app can leave less for the CLI.
• Cloud tasks cost more. Handing a job to the web agent is convenient, and the pricing page warns it may use more of your allowance than a local run.
• Closed source on one side. You can read and fork Codex CLI. You cannot do either with Claude Code.
• Instructions drift. Two instruction files, AGENTS.md and CLAUDE.md, will disagree within a month unless one points to the other.

Two plans worked through

These use only the prices above.

A solo developer who codes most days. On the $20 tier both vendors give a rolling window of work. If you reach for the agent a few times a day, Codex Plus at $20 or Claude Pro at $17 billed annually is the natural start. If you hit the limit twice a week, the next step is $100 on either side, a five-fold jump for Claude (Max says 5x Pro) and a plan in OpenAI's $100, $200 and $500 ladder. Before you jump, check whether switching to a smaller model for easy jobs stays under the limit. Codex lists GPT-6 Luna at 350 to 3,000 messages per window against 15 to 160 for GPT-6.1 Sol, which shows how much the model choice moves the meter.

A five-person team. Anthropic's average of $150 to $250 per developer per month through the API puts five developers at roughly $750 to $1,250 a month if everyone uses the API. Five Max seats would be $500. Five Pro seats would be $100 to $125 depending on billing. These are list-price sums and say nothing about which team gets more done. Use them to find the floor and ceiling before you ask for budget.

A first week with each

Day one: install, sign in with the plan you already have, and run the agent on a read-only task such as "explain how this repo handles auth." That costs little and tells you how well it reads your code.

Day two and three: give it two small bugs from your tracker. Review each diff as you would a colleague's pull request.

Day four: write the project instructions file. Keep it to the commands to build and test, the folders that matter and the things the agent got wrong on days two and three.

Day five: try the review command or skill on a real branch before you push it.

At the end of the week look at your usage page and note how much of the allowance you consumed. That number is the best input to the plan decision, and neither vendor can give it to you.

Our take

From the documents, these tools are more alike than different: terminal agents with MCP, AGENTS.md, skills, a non-interactive mode, a sandbox and a cloud counterpart. Both vendors sell rolling-window plans at $20, $100 and $200.

Where they differ is license and entry. Codex is Apache-2.0 and gives you a no-cost way in. Claude Code is closed and starts at a paid plan, with the most extensive documented extension system of the two.

So the choice rarely comes down to features. If you pay for one subscription already, the other is a second bill, and the tool bundled with your existing plan is the economical start. Our recommendation is to try the bundled one for two weeks on real work and look at your usage page before adding anything. We have not run a benchmark on either, and this page makes no claim about which writes better code.

Related

• Cursor vs Codex
• Codex vs Gemini CLI
• OpenAI Codex CLI
• Claude Code
• Claude Code vs OpenCode
• How we compare