Codex vs Claude Code: Differences, Pricing, and Which to Pick

Codex is OpenAI's coding agent, and Claude Code is Anthropic's. As of September 30, 2026, both run in your terminal, your IDE, a desktop app, and the cloud, and the first full plan costs $20 a month on each side: ChatGPT Plus for Codex, Claude Pro for Claude Code. The real differences are the models behind each tool, how usage limits are counted, how each one asks before it acts, and which team tools it plugs into, so the better pick is usually the one that matches the account you already pay for and the kind of task you hand over.

Codex vs Claude Code side by side (as of September 30, 2026)

All plan prices are in USD per month and come from OpenAI's Codex pricing page, Anthropic's Claude pricing, and the Claude Code docs.

The old shorthand "Codex runs in the cloud, Claude Code runs locally" no longer holds. Both tools edit files on your machine, both can run tasks remotely, and both can open pull requests. What still separates them is the account, the counting of usage, and the safety model.

Is Claude Code or Codex free?

Codex has a free way in, and Claude Code does not. As of September 30, 2026, OpenAI lists Codex on the ChatGPT Free plan ($0) as a way to "explore Codex capabilities on quick coding tasks," with GPT-6 Luna at Standard speed in the desktop app, subject to rollout. The Go plan ($8 a month) gets the same model in the same app for lightweight coding tasks. Neither includes the web, CLI, IDE, or iOS access that starts at Plus.

Anthropic's Free plan does not include Claude Code at all. The cheapest subscription that does is Pro at $20 a month, or $17 a month when you pay $200 up front for a year, per Claude pricing as of September 30, 2026.

Both tools also run on an API key with no subscription. You pay per token instead of a monthly fee, so "free" only describes the sign-in, not the usage. Claude API list prices as of September 30, 2026 are $2 per million input tokens and $10 per million output tokens for Sonnet 5.5, and $4 and $20 for Opus 5.5. Codex on an API key follows OpenAI's standard API pricing and only offers the models your key can access.

Codex Plus or Claude Pro: what $20 buys and how it runs out

Both cost $20 a month as of September 30, 2026. The useful comparison is what the $20 covers and what happens when you use it up, not whose model is smarter.

ChatGPT Plus with Codex

Plus opens every Codex surface: the web, the CLI, the IDE extension, iOS, and the desktop app. It adds cloud integrations such as automatic code review and Slack, and gives you GPT-6 Sol and GPT-6 Luna.

Usage is counted in local messages per five-hour window, and OpenAI publishes ranges rather than fixed numbers. For Plus as of September 30, 2026:

OpenAI calls these "estimates" and says they "are not fixed message limits." Cloud tasks can use more of the allowance than local messages do, and weekly limits may also apply. Because GPT-6 Luna's range is about 20 times GPT-6 Sol's (3,000 ÷ 150), running routine edits on Luna and saving Sol for harder steps stretches the same $20 much further. If you still select GPT-5.5, note that it retires from ChatGPT and Codex on October 14, 2026, though the API is unaffected.

When you reach the limit, the turn in progress can finish. After that you can buy ChatGPT credits or switch to an API key and keep working locally at standard API rates. Run /status in the CLI to see what remains.

Claude Pro with Claude Code

Pro includes Claude Code in the terminal, the IDE extensions, the desktop app, the web, and the mobile apps. Anthropic counts usage as a per-session allowance plus weekly limits rather than publishing a message count. Max 5x at $100 a month and Max 20x at $200 a month raise the per-session allowance to 5 or 20 times Pro's, and weekly limits still apply, per Anthropic's plan comparison as of September 30, 2026.

When you pass the plan limit on Pro or Max, usage credits let you continue. Run /usage inside Claude Code to see where you stand.

Choosing between the two $20 plans

If you already pay for ChatGPT Plus, you already have Codex, and the same is true of Claude Pro and Claude Code. Paying for the second one mostly makes sense when you want the other vendor's models, when your team's tools favor it (GitHub @codex versus Slack and Routines), or when one plan's limit keeps cutting your day short. For heavy solo use, the next steps up differ: ChatGPT Pro starts at $100 a month and currently has no five-hour limit, while Claude Max starts at $100 a month for five times Pro's per-session allowance, with weekly limits on both Claude plans.

How each one asks before it acts

Codex and Claude Code both start in a middle setting that edits your project but stops before riskier moves. They describe it differently.

Codex separates two controls in its sandboxing docs. The sandbox mode decides what it can touch:

• read-only inspects files but cannot edit or run commands without approval.
• workspace-write, the local default, reads and edits inside the workspace and runs routine local commands inside that boundary.
• danger-full-access removes the sandbox and its filesystem and network boundaries.

The approval policy decides when it asks: on-request (the local default) or never.

Claude Code uses a single setting, the permission mode:

• default, shown as Manual, reads without asking and asks for everything else.
• acceptEdits also allows file edits and common filesystem commands.
• plan reads and explores before any change.
• auto allows everything, with a classifier running background safety checks. It is the starting mode for interactive terminal and VS Code sessions from v2.1.283.
• dontAsk runs only pre-approved tools and is meant for locked-down CI.
• bypassPermissions is labeled for "isolated containers and VMs only."

Deny rules in Claude Code block in every mode. For both tools, danger-full-access and bypassPermissions belong only inside an isolated container or VM, never on a laptop that holds your SSH keys and cloud credentials.

Which one to pick: decision rules

Pick by the shape of the task and the account you already have, not by a benchmark.

• If the task is written down, has tests, and ends in a pull request you will review, either tool can take it. Choose Codex if your team works in ChatGPT and assigns work with @codex on GitHub. Choose Claude Code if your team works in Claude and uses Slack or scheduled Routines.
• If the requirement is still moving and you want to steer each step live, choose the tool whose safety setting you trust in your own repo. Codex gives you a hard workspace boundary (workspace-write) and asks when it needs to cross it (on-request). Claude Code lets you start in plan mode, then switch to acceptEdits once the approach is settled.
• If the agent runs in locked-down CI, use Codex in read-only mode or on an API key, or Claude Code in dontAsk mode with pre-approved tools.
• If you use Amazon Bedrock, Google Cloud, or Microsoft Foundry and cannot add another vendor account, Claude Code runs through those providers. Codex on an API key needs an OpenAI API account and loses the cloud features.
• If budget is the only constraint, Codex Free or Go lets you try it in the desktop app for $0 or $8, subject to rollout. Claude Code's cheapest subscription is Pro at $17–20 a month; below that, you pay per token on an API key.

What it costs a team

For companies, the bill depends on whether you buy seats or pay per token.

Claude Team seats cost $25 a month for a standard seat ($20 billed annually) and $125 for a premium seat ($100 annually) as of September 30, 2026, and every paid Claude plan includes Claude Code. A Team or Enterprise seat's allowance resets on a rolling five-hour window and a weekly window, shared with Claude chat and Cowork. ChatGPT Business, which includes Codex, is $20 per user per month billed annually on OpenAI's pricing page as of the same date.

If you pay for Claude Code by API tokens instead, Anthropic's cost guide gives averages from enterprise deployments: about $13 per developer per active day and $150–250 per developer per month, with 90% of users staying below $30 per active day. Those figures describe API-billed usage, not what a Pro subscriber pays. A rough monthly budget follows from them:

• Average: developers × active days × $13. For 5 developers working 20 days, 5 × 20 × $13 = $1,300.
• Ceiling for most users: developers × active days × $30. The same team gives 5 × 20 × $30 = $3,000.

OpenAI publishes no equivalent per-developer average for Codex. Its five-hour ranges are estimates, not fixed limits, so budget a Codex rollout from a pilot week rather than from the table.

How to use Codex and Claude Code together

Yes, you can run both on the same repository. It works when each tool has a separate job and you stand between them.

1. Keep one AGENTS.md at the root with your build commands, test commands, and conventions. Codex reads it, and Claude Code can read it on its own or alongside a short CLAUDE.md for Claude-specific notes, so the rules never drift apart.
2. Give each agent its own branch or git worktree. Never let both edit the same working tree at once.
3. Let one agent implement a scoped task and stop at the first diff.
4. Read the diff yourself: intent, tests, changed files, and anything that touches secrets.
5. Hand the diff to the other agent for review only. In Codex, /review runs a local code review; in Claude Code, ask for a review in plan mode so it cannot edit.
6. Merge only after your final check. Neither agent merges its own work.

Skip the second agent when the task has no acceptance criteria or the first agent made broad changes you have not read yet. A second pass on an unclear task adds more changes before it adds confidence.

FAQ

Can Codex use Claude models, or Claude Code use GPT models?

Not according to either vendor. As of September 30, 2026, OpenAI's Codex docs list only OpenAI models, and Anthropic's Claude Code docs list only Claude models, reached directly or through Amazon Bedrock, Google Cloud, or Microsoft Foundry. Third-party wrappers that swap models in are outside both vendors' documentation.

What is Claude's equivalent of Codex?

Claude Code. It is Anthropic's coding agent, included in every paid Claude plan from Pro upward, and it covers the same ground as Codex: editing a repository, running commands, and opening pull requests.

Is Claude Code still terminal-only?

No. It started in the terminal, but as of September 30, 2026 it also runs in VS Code, Cursor, JetBrains, a desktop app, the web at claude.ai/code, and the Claude mobile apps, and it can run on a schedule through Routines.

Does Claude Code read AGENTS.md?

Yes. Anthropic's overview says that if a repository already has an AGENTS.md for other coding agents, Claude Code can read it on its own or alongside CLAUDE.md. That makes a single shared instruction file practical when a team uses both tools.

Which is better at writing code?

Neither vendor publishes a head-to-head comparison, and the models change often. Give both the same small task from your own backlog, with the same tests and the same review standard, and keep the one whose diff you would merge with fewer corrections.