Among AI coding assistants, two tools resemble each other more than any others: Claude Code, Anthropic’s agent, and Codex, OpenAI’s. Both are agents you delegate a task to (read the project, edit files, run the tests) rather than completion tools that suggest the next line. Both run in the terminal, in VS Code, in a desktop app and in the cloud. And, as of October 2026, their price lists are almost identical, down to the dollar.
For a small business, then, the question isn’t “which one is cheaper?” It’s twofold: which subscription the tool comes with (Claude or ChatGPT, which the company may already pay for), and who says yes on the developer’s behalf when the agent works on its own. This comparison was built from both vendors’ pricing pages and documentation, checked on October 15, 2026, the day after GPT-5.5 was removed from Codex; US prices were rechecked on November 2, 2026. It is the fourth part of our “Coding with AI” series, after Claude Code vs Cursor, Cursor vs GitHub Copilot and Lovable vs Bolt.
For an overview of the six tools in the series (agents, AI-assisted editors, app builders), team budgets and the rules to set, read our guide Coding with AI in a small business: which tool to choose.
The 30-second answer
Choose Claude Code if…
- your company already uses Claude (or plans to) for its non-technical teams;
- your developers want an agent that works on its own through long tasks, with a second model that filters risky actions;
- you want up to $500 a month in pooled API credits included with your team plan (as of November 2026, both vendors offer Standard seats at $20 and Premium seats at $100, billed annually);
- JetBrains, Chrome or Slack are part of your daily tools.
Choose Codex if…
- your company already pays for ChatGPT Plus or Business: Codex is included, with no extra subscription;
- you prefer a simple technical barrier: an OS-level sandbox with no network access by default;
- you want an open-source command-line tool (Apache 2.0 license) that you can audit;
- you want automatic review of GitHub pull requests with a simple “@Codex”.
The comparison, line by line
| Criterion | Claude Code | Codex |
|---|---|---|
| Vendor | Anthropic (United States) | OpenAI (United States) |
| Where it runs | Terminal, VS Code, JetBrains, Claude desktop app, browser (cloud), mobile, Slack | Terminal, VS Code and compatible editors, ChatGPT desktop app, cloud, iOS, Slack |
| Subscriptions that include it | Claude Pro, Max, Team, Enterprise | ChatGPT Plus, Pro, Business, Enterprise (Free and Go: limited access in the desktop app) |
| Individual plan | Pro: $20/month ($17 billed annually); Max: $100; Ultra: $200 | Plus: $20/month; Pro: $100, $200 or $500 |
| Team plan | Team: Standard seat $20 (annual) or $25 (monthly); Premium $100 or $125; 2 to 150 people | Business: Standard seat $20 per user (annual, 2-user minimum) or $25 (monthly); Premium $100 or $125; up to 200 seats |
| Large enterprise | Enterprise: $20/seat/month + usage at API rates, 20-seat minimum | Enterprise: pricing not published (contact sales) |
| Quota | Rolling 5-hour window + weekly cap, shared with chat and Cowork | 5-hour window (Plus, Business Standard) + possible weekly caps, shared with ChatGPT Work |
| Default safeguard | “Auto” mode: a second model reviews actions before they run | OS-level sandbox: writes limited to the project, network off, approval on request |
| Instructions file | CLAUDE.md (also reads AGENTS.md) | AGENTS.md |
| Source code | Proprietary | Open-source command-line tool (Apache 2.0, written in Rust) |
| Training on your data | Never by default on Team, Enterprise and API; an account setting on Pro and Max | Never by default on Business and Enterprise; an account setting on Plus and Pro |
Prices in US dollars, excluding tax, recorded on October 15, 2026 and rechecked on November 2, 2026 on claude.com/pricing, learn.chatgpt.com/docs/pricing and OpenAI’s ChatGPT Business help center.
Same price down to the dollar: the difference lies elsewhere
Put the two price lists side by side: $20 for the entry-level individual plan, $20 per person on a team plan with an annual commitment, $25 without commitment, $100 and $200 for heavy users. The table settles nothing, and that’s by design: each vendor matches the other. What differs is what the plan actually buys, and neither puts it in firm numbers.
At OpenAI, the Codex documentation gives ranges of “local messages” per five-hour period, identical on Plus and Business: 15 to 160 messages with GPT-6.1 Sol, the recommended model; 5 to 45 with GPT-6 Astra, the most powerful; 350 to 3,000 with GPT-6 Luna, the lightest. The vendor states that these are “not fixed limits,” that weekly caps may also apply, and that tasks run in the cloud use more. A range that varies tenfold doesn’t let you budget; it only tells you that everything depends on the size of the task.
At Anthropic, a Team member’s quota resets on a rolling five-hour window and on a weekly window; the Premium seat gives “5x more usage” than Standard, the Max plan “5x” Pro usage and the Ultra plan “10x” Pro usage per week (as of November 2026). It’s a relative scale, with no published absolute value.
Two things they have in common matter more than these figures. First, the quota is shared with the rest of the assistant: at Anthropic with chat and Cowork, at OpenAI with ChatGPT Work (and, on Plus and Pro, with ChatGPT for Excel). A developer who also spends the day in the chat draws on the same budget. Second, overage is paid in credits on both sides: ChatGPT credits on Plus and Pro, workspace credits on Business; extra usage on Pro, Max and Team at Anthropic. Either way, the admin should set a cap as soon as the account is opened.
One difference, however, is new: since the latest update to its pricing page, Anthropic includes $100 to $200 a month in API credits in the Max ($100) and Ultra ($200) plans, and up to $500 a month in pooled API credits for a Team plan. OpenAI has no equivalent on its Codex pricing; in return, its Pro plan (like the Business Premium seat) currently has no five-hour limit (Anthropic’s Ultra plan now offers the same for verified accounts).
Who says yes when the developer isn’t there?
A coding agent is only useful if it can chain actions without asking permission at every line. But every unapproved action is a risk: a file deleted outside the project, a command that sends data over the internet, a secret read by mistake. The two vendors made different choices for their default setting.
Claude Code judges the action. In recent versions, a session in the terminal or in VS Code starts in “auto” mode by default: instead of asking for your approval, a second model, the classifier, reviews each action and blocks the ones it considers risky. Manual mode, which asks before almost every edit or command, and “plan” mode, which forbids any edit until you’ve approved a plan, are still available. A sandbox for commands also exists, but you have to turn it on.
Codex sets the boundaries. By default, the agent works in a sandbox enforced by the operating system: it can read, edit and run freely inside the project folder, but network access is off, and it must ask for your approval to write elsewhere or connect to the internet. OpenAI also offers model-based “automatic approval review,” but as a complement to the barrier, not a replacement.
For a small business, the difference is concrete. An OS-level barrier is verifiable: you know what it prevents, regardless of how good the model’s judgment is. A classifier is more flexible: it lets a legitimate dependency install through without interrupting the developer, but its judgment isn’t guaranteed. Neither approach exempts you from the basic rule: the agent works on a copy of the code, never with production access, and its changes go through review before merging. That’s also what we recommend in our getting-started guide to Claude Code.
What a team of 3 developers costs
As with the other tools in the series, a small business that hands its code to an agent should use a team plan: that’s what brings centralized administration, single sign-on (SSO) and no training by default. Ballpark figures before any overage, excluding tax, as of November 2026.
| Situation | Claude Code | Codex |
|---|---|---|
| 1 developer, individual plan | Pro: $20/month | Plus: $20/month |
| 3 developers, annual commitment | 3 Standard seats × $20 = $60/month | 3 Business users × $20 = $60/month |
| 3 developers, no commitment | 3 × $25 = $75/month | 3 × $25 = $75/month |
| Including 1 heavy user (annual) | 2 × $20 + 1 Premium seat at $100 = $140/month | 2 × $20 + 1 Premium seat at $100 = $140/month (or 3 × $20 + workspace credits as consumed) |
| What the team gets on top | Claude for the whole team (chat, Cowork, Design…), up to $500/month in pooled API credits | ChatGPT for the whole team (chat, Work…), more powerful virtual machines for the cloud |
So the real calculation isn’t “Claude Code or Codex,” but “which general-purpose assistant is the company already paying for?”. If your sales team and leadership have ChatGPT Business licenses, adding developers to the same workspace costs only their seats, and Codex comes with them. If the company chose Claude Team, Claude Code is already paid for. Taking both means doubling the bill for a very similar service. To compare the general-purpose assistants themselves, read our comparison of ChatGPT, Copilot and Claude for business.
As of November 2026, the structure is the same at both vendors. Anthropic and OpenAI both let you mix seat types: you pay for a Premium seat ($100 a month billed annually, $125 monthly, at both vendors) only for the developer who runs the agent all day. Beyond the included allowance, heavy users draw on credits (workspace credits at OpenAI, extra usage at Anthropic), which the admin buys and caps.
Models: yesterday’s comparison no longer holds
On October 14, 2026 at 7 p.m. UTC, GPT-5.5 was removed from Codex for all users signed in with their ChatGPT account. OpenAI asks users to replace it with GPT-6.1 Sol on Plus, Pro, Business and Enterprise, and with GPT-6 Luna on Free and Go, including in admin configurations, custom agents, scheduled tasks and scripts. Sessions opened with an API key are not affected. In other words, most “Claude Code vs Codex” benchmarks published before that date compare a model that is no longer offered in the subscription.
On Anthropic’s side, Claude Code gives access to the current Claude models, with a choice of model and reasoning effort within the session. The same logic applies at both vendors: the more powerful the model, the faster the quota runs out. OpenAI’s documentation spells it out: “Fast” mode uses 2.5 times more included quota, and “Ultrafast” mode 8 times more.
The practical consequence: don’t choose based on a model leaderboard. They change every two months, and each vendor takes the lead in turn. Choose based on what doesn’t change: the subscription you already pay for, the safeguard that fits your risk level, and how the tool integrates with your repositories.
Data, privacy and administration
On the question that worries business owners most, the two vendors have aligned their rules, and they depend on the plan, not the tool.
- Team and enterprise plans (Claude Team and Enterprise, ChatGPT Business and Enterprise): no model training on your content by default. Same rule for API key usage.
- Individual plans (Claude Pro and Max, ChatGPT Plus and Pro): it’s an account setting. Anthropic’s pricing page lists it as “opt-out”; at OpenAI, it’s managed in the “Data controls” settings. At Anthropic, conversations from users who allow training are kept for 5 years, versus 30 days otherwise.
- Large enterprise: Claude Enterprise adds custom retention controls, and zero retention for Claude Code is granted case by case by the sales team; ChatGPT Enterprise adds retention and data residency controls, plus a compliance API for audits.
Two details matter for an IT department. Claude Code stores session history locally in plain text, for 30 days by default, in the user’s folder: factor this into your endpoint policy. Codex’s command-line tool is open source (Apache 2.0): your security team can read what it does, even though the models stay with OpenAI. Claude Code, by contrast, is proprietary software.
Finally, whatever the tool, a small business should require a team plan for anyone who touches company code, rather than letting each developer use a personal subscription. We cover this in detail in our article on Claude at work: data and security.
Integration: GitHub, Slack and instruction files
Both agents can work in the cloud on a GitHub repository and hand back a pull request for review. Codex stands out for automatic code review: just mention “@Codex” in a pull request, or turn on automatic review for a repository. Note that these cloud features are only available with a ChatGPT account: in API key mode, Codex works in the command line, the SDK and the editor, but without GitHub review or Slack. Claude Code stands out for the range of its entry points: a JetBrains extension (IntelliJ, PyCharm, WebStorm…), Chrome control to test a web application, and sessions you can move between the terminal and the cloud.
For project instructions, Codex reads an AGENTS.md file at the root of the repository, and Claude Code a CLAUDE.mdfile, but its documentation states that it also supports AGENTS.md. A team that can’t decide between the two, or lets each person choose, should therefore write its rules in AGENTS.md: code conventions, test commands, off-limits folders. Switching tools then becomes painless. Our methods for writing instructions for Claude apply to both.
What this table doesn’t tell you
Identical prices prove nothing
When two price lists match down to the dollar, the comparison shifts to what the plan buys. And neither publishes a firm volume: a range of 15 to 160 messages at OpenAI, a “5x more” scale at Anthropic. The only judge is the usage dashboard after a two-week trial.
The model tested yesterday no longer exists
GPT-5.5 left Codex on October 14, 2026. A comparison, benchmark or demo made before that date describes a product that is no longer sold. Always check when a comparison’s data was collected, including this one.
How to decide in two weeks
- Start from the subscription you already have. If the company already pays for ChatGPT Business or Claude Team, start with the included agent: the trial costs nothing extra.
- Write an AGENTS.md. One page: how to run the tests, the conventions, what the agent must never touch. It will work for both tools.
- Give both agents the same task. A real task from your repository, on a dedicated branch, for example a documented bug fix or a small module with its tests. Compare review time, not just the result.
- Set the safeguard and read the usage. Choose the level of autonomy based on how sensitive the repository is, then check the usage dashboard after a week (the
/statuscommand in Codex,/usagein Claude Code). - Move to a team plan with a cap. Seats only for active developers, a spending cap set by the admin, personal subscriptions banned for company code.
If you’d like support with this choice, from setting usage rules to training your teams, explore our consulting services in AI strategy.
Frequently asked questions
What is the difference between Claude Code and Codex?
They are two coding agents you delegate tasks to: Claude Code is made by Anthropic and included in Claude subscriptions; Codex is made by OpenAI and included in ChatGPT subscriptions. Both run in the terminal, in VS Code, on the desktop and in the cloud. The main difference is the default safeguard: a second model that filters actions in Claude Code, an OS-level sandbox with no network in Codex.
Which one is cheaper for a small team?
Neither: as of November 2, 2026, both cost $20 per person per month on a team plan with an annual commitment (Claude Team Standard, ChatGPT Business Standard) and $25 without commitment, excluding tax. The cheapest one is the one included in the assistant your company already pays for.
Is Codex included in ChatGPT Plus?
Yes. Codex is included in ChatGPT Plus, Pro, Business and Enterprise, in the terminal, the editor, the desktop app and the cloud. The Free and Go plans give limited access in the desktop app. Codex’s quota is shared with ChatGPT Work.
Is my code used to train Claude or OpenAI’s models?
Not by default on team or enterprise plans (Claude Team and Enterprise, ChatGPT Business and Enterprise), nor via API key. On individual plans, it’s an account setting each user can turn off. A small business should keep company code on team plans only.
Can you use both on the same project?
Yes. Codex reads the AGENTS.md file and Claude Code supports it too, in addition to its CLAUDE.md. Writing the project rules in AGENTS.md lets you switch from one tool to the other without rewriting the instructions. Paying for both subscriptions for the same team, however, is rarely justified.
In short
Claude Code and Codex are the two most similar coding agents on the market, sold at the same price: $20 for an individual plan, $20 per person on a team plan with an annual commitment (as of November 2026). The choice comes down to three stable criteria: the assistant your company already pays for (Claude or ChatGPT), the default safeguard (a model that judges at Anthropic, an OS-level barrier at OpenAI) and integration with your tools (JetBrains and Chrome on one side, “@Codex” GitHub review and an open-source tool on the other). Model leaderboards, on the other hand, change every two months, as the removal of GPT-5.5 on October 14 showed. To put these tools on a sound footing in your company, explore our consulting services in AI strategy or contact us.