Claude Code vs Codex: Which Should You Choose? Pricing, Limits, Where They Run, and Features Compared

Anthropic's Claude Code and OpenAI's Codex are both coding agents: they read your repository, write code, run commands, and carry the work through to a finished set of changes on their own. When we re-read both companies' official documentation in the original on September 22, 2026, we found that the places they run (terminal, IDE, desktop app, cloud, phone) and the core features (subagents, MCP, hooks, skills) now largely match.

So the real differences today are less about what each tool can do and more about the conditions under which you can use it. Can you try it for free? Does the sandbox work on Windows? Which operating systems and plans get computer use? Can you run GitHub PR reviews on a personal plan? And underneath all of that: Claude Code runs on Claude models, and Codex runs on OpenAI models.

1. The short answer: the gap is in access conditions, not features

Here is the whole picture on one page. Sections 2 onward check each condition and its source one by one.

None of the differences in the lower half of the table are features that only one side has. They are the same features offered under different conditions. That means you can decide by checking, in order: (1) which subscription you already pay for, (2) which models you want to use, and (3) whether your setup (OS, plan, region) meets the conditions. Answers by scenario are in Section 8.

2. Where they run: terminal, IDE, desktop, cloud, phone

The old split of "Claude Code for interactive local work, Codex for handing tasks off to the cloud" no longer holds. Both run locally and in the cloud. What differs are the details of each entry point.

Sources: Claude Code "Advanced setup", VS Code, Use Claude Code in the cloud, mobile / Codex CLI, IDE extension, Codex cloud, Remote, What's new (all checked September 22, 2026)

How they differ in the cloud

Both share the main benefit of the cloud: the work keeps going after you close your laptop. Beyond that, they differ in three ways.

Sources: Claude Code "Configure cloud environments" / Codex "Cloud environments", Codex Cloud (Legacy), Codex "Pricing", Workspace model availability (Codex environments, network access, and integrations rechecked October 8, 2026)

Claude Code also has a Projects feature that hands out multiple cloud sessions from a single conversation and pulls their progress together. Codex starts a separate working environment for each task from a published environment. Codex Cloud (Legacy) continues to handle Code Review, Linear, and GitHub integrations. Delegation to new Cloud environments from Slack and Microsoft Teams is for Enterprise workspaces with Cloud delegation enabled.

3. Pricing and usage limits: free tiers, 5-hour windows, weekly caps, overages

In both cases, the coding agent comes bundled with a chat subscription: Claude Code with Claude plans, and Codex with ChatGPT plans.

Sources: Claude pricing page, Max plan help article, Claude Code "Advanced setup" / Codex "Pricing", OpenAI Help "About ChatGPT Pro tiers", X post by OpenAI staffer Tibo, OpenAI "DevDay 2026 Recap", OpenAI Help "Using Codex with your ChatGPT plan" (checked September 22, 2026; the three OpenAI pages rechecked September 26, and the Pro tiers and multipliers rechecked October 5)

The $20, $100, and $200 tiers are shared, and ChatGPT adds a $500 Pro tier on top. The clear difference is at the bottom end. OpenAI's help center says Codex is included in every ChatGPT plan, Free and Go included, while Claude Code's official documentation says Claude Code is not included in the free claude.ai plan. If you want to try one without paying anything, Codex is the only way in.

Both have a 5-hour window and a weekly cap; they differ in how you pay for overages

In both, your allowance goes down with the weight of the work (how much is read, how much is written, which model is used), both have a 5-hour window and a weekly cap, and both let you pay extra to keep going once you hit them. What differs is the unit for the overage.

They also publish numbers differently. OpenAI gives per-model guideline figures for messages per 5 hours, but the ranges are wide, with the high and low ends nearly 10x apart even for the same model and plan, and OpenAI itself notes that these are not fixed limits. Anthropic gives no counts, saying the number of messages you can send varies with message length, attachments, conversation length, and model. With either one, you won't know how much headroom you really have until you've run it for a week or two with your own workload.

What Claude Code shows when you hit the limit, and what to do about it, is covered in our article on "You've hit your session limit"; to find which session is eating your allowance, see how to view usage by session. Codex's limits are also shared with ChatGPT Work, so if you use Work heavily, see our explainer on ChatGPT Work too.

4. How the work runs: mid-task steering, parallelism, sandboxes and permissions

Can you change direction mid-task?

Yes. Both let you send additional instructions while they work. In the terminal, Claude Code queues what you type during a task and passes it on at a natural break; you can also press Esc to stop and redirect. Codex CLI can likewise inject instructions into a turn that is already running.

The same goes for the cloud. Claude Code's documentation says sessions keep running after you close your laptop and that you can check in and steer from any device; Codex's documentation says you can continue in the same chat to add context, change constraints, or ask for a progress summary. Both can change course partway through cloud tasks (Claude Code "Interactive mode", Codex "Long-running work").

Ways to run things in parallel

Sources: Claude Code "agent teams", Desktop, cloud / Codex "Subagents", Models, Worktrees

For when to use which on the Claude Code side, see subagents vs. agent teams and our article on agent view. In Codex, subagents start when you ask for them or when AGENTS.md tells it to split the work; how to run them in parallel in both tools is covered in Codex and Claude Code Subagents: How to Run Them in Parallel. With both companies, the more you parallelize, the faster you burn through your limits.

Sandboxes and permissions: the defaults differ most

This is where the two designs diverge most clearly.

Sources: Claude Code "Choose a permission mode", Sandboxing / Codex "Sandbox", Permissions, Windows sandbox

In practice: if you want to stay on plain Windows and let the agent work inside an OS-level fence, Codex is the more straightforward choice. Claude Code runs on native Windows too, but for the same kind of fence you run it inside WSL2. Conversely, Claude Code's auto mode is a way to cut down on confirmation prompts, not a mechanism that fences off files and network access at the OS level. For important repositories, it's safer to use /sandbox alongside it.

5. Instruction files and extensions: CLAUDE.md, AGENTS.md, MCP

The file that passes your project's conventions (build commands, how to run tests, coding style) to the agent is CLAUDE.md for Claude Code and AGENTS.md for Codex. AGENTS.md is an open format maintained by the Agentic AI Foundation under the Linux Foundation, and tools other than Codex read it too (agents.md).

There is one important change. From v2.1.277, if there is no CLAUDE.md, .claude/CLAUDE.md, or CLAUDE.local.md in the working folder or above it, Claude Code reads AGENTS.md as is. If any of those files exists, only the CLAUDE.md side is read by default, so if you use both, add one line to CLAUDE.md to import it (Claude Code "memory").

@AGENTS.md

## Claude Code only
- Delegate research to subagents

The above is an example CLAUDE.md. The first line imports AGENTS.md; below it, write anything you want to apply only to Claude Code.

Codex reads ~/.codex/AGENTS.md (shared across all repositories) and then every AGENTS.md from the repository root down to the working folder, concatenated in order. Once the total reaches the default of 32 KiB, nothing more is added, so in large repositories you split them up per folder. By setting project_doc_fallback_filenames, you can also have it read a file under another name in folders that have no AGENTS.md (Codex "AGENTS.md").

If you've written rules and they're not being followed, why AI ignores your rules and how to fix it covers what to check in each tool.

MCP, hooks, skills, plugins

MCP for connecting to external tools and data, hooks that run scripts at set points, skills for reusable procedures, and plugins that bundle them together: both tools have all of these. There's little reason to choose between them on this basis. What's worth remembering is how they relate to your limits. On its pricing page, OpenAI advises that each MCP server adds context and uses more of your limits, so you should disable servers you aren't using. Claude Code's documentation, too, makes deferred loading of tool definitions (loading them only when needed) the default and recommends disabling unused servers with /mcp (Claude Code "Manage costs").

6. Computer use: which OS, plan, and region

Computer Use, which lets the agent operate your PC's screen, is available in both Claude Code and Codex. Both look at the screen and click and type to drive native apps, simulators, and GUI-only tools that commands and APIs can't reach. The difference is in availability.

Sources: Claude Code "computer use", Desktop / Codex "Computer Use", Pricing (feature table)

Both are designed on the principle of reaching for precise tools first and treating screen control as a last resort. Claude Code tries MCP, the shell, and the Chrome integration, in that order, before moving to screen control, and Codex advises using its built-in browser first if you're working on a local web app. If you want computer use under a company Team or Enterprise contract, Claude Code currently can't do it, which is worth factoring into your decision.

7. Models and performance: choosing a tool means choosing a model

Claude Code runs only on Claude models, and Codex only on OpenAI models (Claude Code can also use Claude through Amazon Bedrock, Google Cloud, and Microsoft Foundry; Codex can also use OpenAI models through Amazon Bedrock and Google Cloud (Vertex AI)). In other words, "which tool?" is nearly the same question as "which company's models?"

Model names turn over every few months, so none of the recommendations in this article depend on them. The two cards above are the only dated information, and we'll swap them out when the models change. We also track each company's current models in our list of major models.

Why we don't include a benchmark comparison table

This article doesn't include a benchmark table such as SWE-bench Verified. Like our Claude Opus 5.5 and GPT-6 Astra articles, we only put side by side figures measured by the same publisher with the same yardstick. There are three reasons.

If you want to check the numbers, look at each company's release post and see whether the same test lists scores for both. A more reliable approach is to compare them on your own repository using the steps in Section 9.

8. Which one to pick: nine scenarios

Here's how the conditions above apply to common situations. If more than one fits, give priority to the one higher up.

Sources: Claude Code "Code Review", GitHub Actions, Claude Code's LICENSE.md / Codex "Pricing", Codex "Open Source", OpenAI Help "About ChatGPT Pro tiers"

9. Using both, switching, and testing them on your own repo

Using both costs $40 a month with Claude Pro and ChatGPT Plus. You can also carry on with one when the other runs out of allowance. If you use both, consolidating your instructions into AGENTS.md saves you from maintaining your conventions in two places.

If you're moving from Claude Code to Codex, Codex has an import feature. Type /import in Codex CLI and choose Claude Code to bring over your settings, project files, and recent conversations (up to 50 from the last 30 days). The ChatGPT desktop app can import from Claude Code, Claude Cowork, and Cursor. Your settings in the original tool are left unchanged (Codex "Import from another agent").

A 30-minute way to compare them

What tells you more than any benchmark is how each one handles the same tasks on your own repository. With the steps below, one trial takes about 30 minutes.

If this comparison shows no difference, either tool will do. In that case, the fastest way to decide is on pricing and setup conditions (Sections 3–6).

For dedicated security scans, see our guide to Codex Security setup, pricing, and interpreting results.

To keep completion criteria in view while Codex investigates, fixes, and verifies, see how to use Codex /goal. The guide covers desktop and CLI controls, stopping conditions, and the distinction between token budgets and pricing.

FAQ

Q1. Can I use Codex for free? What about Claude Code?

Codex, yes. It's included in ChatGPT Free ($0) and Go ($8); the limits are just small. Claude Code isn't included in the free Claude plan; you need Pro or higher, or pay-as-you-go billing with a Claude Console API key. The evidence is summarized in Section 3.

Q2. Isn't Claude Code local-only?

No. It can also run sessions on cloud VMs managed by Anthropic (Pro, Max, Team, and some Enterprise seats; left research preview on September 23, 2026). The work continues after you close your laptop, and you can add instructions while it runs. Entry points and conditions are in the table in Section 2.

Q3. Will my code be used for training?

On individual plans, it depends on your settings for both. Anthropic says that for Free, Pro, and Max, users choose whether their data may be used for training, and if they opt in, that includes Claude Code usage; under commercial terms such as Team, Enterprise, and the API, data isn't used for training by default. OpenAI says your ChatGPT training data setting also applies to content processed by Codex, and that Business, Enterprise, and API inputs and outputs aren't used for training by default (Claude Code "Data usage", OpenAI Help). Note that even when running locally, whatever you pass to the model is sent over the network to that company. For the steps to opt out on an individual plan, see "How to opt out of Claude Code training."

For OpenAI products, see ChatGPT and Codex training settings and confidential data risks. The guide also covers the separate setting for training on Codex environments and what to do after entering confidential information.

Q4. Is it OK to use both on the same repository?

Yes. There are two things to watch. The first is instruction files: make AGENTS.md the source of truth and have Claude Code import it with @AGENTS.md, and they won't drift apart. The second is simultaneous editing: if two agents write to the same working folder at the same time, they may overwrite each other's changes. If you run them in parallel, give each its own working folder with worktrees.

Q5. Codex says "thread not found."

That doesn't necessarily mean the conversation history itself is gone. We've written up the steps to check before deleting anything in how to fix "thread not found" in Codex.

Sources

• Claude Code documentation: Platforms and integrations / Advanced setup / Use Claude Code in the cloud / Cloud environments / Permission modes / Sandboxing / Memory (CLAUDE.md, AGENTS.md) / Computer use / Model configuration / Code Review / Data usage
• Anthropic: Claude pricing page / Max plan / Using Claude Code with Pro and Max / Usage credits
• Codex documentation: Pricing / Codex CLI / IDE extension / Codex cloud / Sandbox / AGENTS.md / Computer Use / Models / Open Source / Import from another agent / What's new
• OpenAI: Using Codex with your ChatGPT plan / About ChatGPT Pro tiers / Deprecations
• Licenses: openai/codex (Apache-2.0) / anthropics/claude-code LICENSE.md / agents.md

All of the above were checked against the original text of the official documentation retrieved on September 22, 2026. Models and defaults, cloud availability, and the Codex pricing page were retrieved again and rechecked on September 26. The ChatGPT Pro tiers and the Codex cloud model were rechecked against the original text on October 5. Pricing, plans, and availability can change, so check each company's pages before you sign up.

Related: Claude Fable 5 for coding.

Related: What Is ChatGPT Work? The GPT-6 Astra work agent — OpenAI's work agent that shares usage limits with Codex.