Claude vs Codex 2026: Claude Code or OpenAI Codex

Almost every developer chat we sit in this autumn asks the same thing: Claude vs Codex, which one should the team pay for. The question is fair, because both agents now write code, review pull requests and work alone for an hour without anyone watching, and both start at about €18 a month.

We use Claude Code every day at satori., which makes us biased, so this comparison leans on the vendors' own documentation and on independent test results checked on 9 November 2026. You get what each tool does differently, what it costs in euros, how it handles European data and where Codex is simply the better buy.

Claude vs Codex in brief, checked on 9 November 2026

---

Claude vs Codex: the short answer

One short answer covers most teams, and the rest of this article explains how we got there.

• A large codebase where the agent must reason across many files and your own systems: Claude Code, which connects to internal tools through MCP and holds the top independent coding scores right now.
• A company already on ChatGPT Business that wants one bill: Codex, because every Business seat includes it, and our managed ChatGPT Business service invoices those seats in euros.
• Many small, well-defined tasks sent off in parallel: a draw, since both agents now run cloud tasks side by side and return branches for review.
• A tight budget for heavy API use: Codex with GPT-6.1 Sol, which costs the same per token as Sonnet 5.5 but scores lower and runs cheaper per task.
• Still unsure: give both the same real task for a week, which remains the only benchmark that counts.

---

What is the difference between Claude Code and OpenAI Codex?

Claude Code is Anthropic's agentic coding tool, meaning a program that reads your codebase, edits files, runs commands and keeps working until a check passes.¹ It runs in the terminal, in VS Code and JetBrains, in a desktop app and in the browser, and a plain-English tour for non-developers sits in what Claude Code is. Codex is OpenAI's answer, and it now lives inside ChatGPT: the desktop, web and mobile apps, the Codex CLI, an IDE extension and Codex Cloud all reach the same agent.²

A year ago the neat summary was that Claude Code ran on your laptop while Codex ran in OpenAI's cloud. That line has gone. The Codex CLI works against your local repository and runs the tools already installed on your machine.³ Claude Code, for its part, now runs cloud sessions in isolated virtual machines that Anthropic manages, and they keep going after you close the laptop.⁴

The difference that remains is how each one treats safety by default. Local Codex runs in a sandbox with network access switched off and write permission limited to the active workspace, so the agent cannot reach the internet unless you allow it.⁵ Claude Code asks for your approval before it edits a file or runs a shell command, apart from a short list of read-only commands.⁶ A team can save those approval rules in the repository, so every developer starts from the same settings. Both approaches work, as long as nobody on the team approves every prompt without reading it.

---

Which models sit behind Claude Code and Codex?

Claude Code uses Anthropic's current line-up: Haiku 5.5, Sonnet 5.5 and Opus 5.5, with Fable 5.1 above them for the hardest reasoning.⁷ Claude Opus 5.5, released on 22 September 2026, is the usual starting point, and Sonnet 5.5 costs US$2 per million input tokens and US$10 per million output tokens through the API, half the Opus price.⁷

Codex runs on OpenAI's GPT-6 family. OpenAI describes GPT-6 Astra as its most capable model for complex work and GPT-6.1 Sol as near Astra at a lower cost.⁸ GPT-6.1 Terra is the fast model for clear writing and computer use, while GPT-6 Luna is the most efficient choice for focused, high-volume coding. The older GPT-5.3-Codex that many 2026 comparisons still quote has been retired from Codex.⁸ GPT-6.1 Sol costs exactly the same as Sonnet 5.5 on the API, US$2 in and US$10 out per million tokens, with a context window of 1,050,000 tokens.⁹

---

How do Claude and Codex compare on coding tests?

Every launch page shows a chart where the new model wins, so we prefer one independent source. DeepLearning.AI's report on the GPT-6.1 Sol launch summarises the Artificial Analysis results, where Claude Opus 5.5 scores 58 points on the combined Intelligence Index against 53 for GPT-6 Astra and 52 for GPT-6.1 Sol.¹⁰ On the Artificial Analysis Coding Agent Index, which averages three agent coding tests, Claude Sonnet 5.5 and Opus 5.5 hold the top two places.¹⁰

Codex wins clearly on cost per task. The same report puts GPT-6.1 Sol at US$0.72 per task on the Intelligence Index against US$3.26 for Astra.¹⁰ On the coding index, Sol in Codex even beat Astra by one point at under 15 per cent of its cost per task. A team that runs hundreds of small automated jobs a day should take that number seriously, because a few points of quality can matter less than a bill a quarter the size.

Treat every test score as a direction rather than a promise. Your own repository, with its odd build scripts and ten-year-old modules, will look more like a stranger's test than like a vendor's best run, so the one-week trial below is worth more than any chart.

---

What does Claude Code do better than Codex?

Claude Code is built for long, connected sessions in one codebase, where the agent must keep track of why twenty files depend on each other. The independent coding results above back that up today, with Sonnet 5.5 and Opus 5.5 ahead of every GPT-6 model on the agent coding index.¹⁰

Claude Code and MCP: links to your own systems

For European companies the bigger practical difference is reach. Claude Code talks to internal systems, ticket tools, databases and documents through MCP, the Model Context Protocol, an open standard for connecting AI tools to data sources.¹ Codex supports MCP as well, so this is no longer a feature only Claude has, although Claude Code's ecosystem of skills, hooks and plugins is older and larger in our experience.²

Claude Code beyond developers

Claude Code also reaches people who never open a terminal, through the Code tab in the Claude desktop app, and the same Team seat covers Claude Cowork for documents and spreadsheets.¹ A company that wants one AI subscription for both developers and office staff gets it in Claude Team at €18 per user a month on annual billing.¹¹

---

Where does Codex beat Claude Code?

We build our whole company around Claude, which is exactly why we should be clear about where Codex wins.

• A free tier. Codex is included in ChatGPT Free and Go for quick, light coding tasks, while Claude Code needs at least a paid Pro plan.¹²
• One bill with ChatGPT. Every ChatGPT Business seat includes Codex alongside ChatGPT and ChatGPT Work, so a company already on Business adds coding without a new vendor or a new contract.¹²
• Cost per task at volume. GPT-6.1 Sol in Codex ran at under 15 per cent of Astra's cost per task on the coding index, which suits large numbers of routine jobs.¹⁰
• EU data residency on Enterprise. OpenAI's European residency covers the Codex app and the command line on Enterprise and Edu, although Codex on the web stays outside it.¹³
• API keys on Plus. Plus, Business and Enterprise users can run Codex with an API key for scripted jobs such as CI, the pipelines that test code on every change, billed at API rates.¹²

The last point needs a caveat on the Claude side, since Claude Code also runs on an API key and on Amazon Bedrock. Codex simply lets a Plus subscriber mix the two on one account, which is handy for a single developer.

---

What do Claude Code and Codex cost in Europe?

Anthropic shows European visitors its prices in euros before VAT, while OpenAI lists ChatGPT in US dollars, so the euro figures for Codex use the reference rate of 1 EUR to 1.1403 USD.¹¹¹²

At list price the two sit on the same line for most teams. The details change the sums: OpenAI says plan usage will share one limit across ChatGPT and Codex from December 2026, and Claude's Pro and Max limits are already shared between the chat and Claude Code.¹⁴¹⁵ Eight developers on annual team seats cost €1,728 a year on Claude Team and US$1,920, about €1,684, on ChatGPT Business, so the subscription the company already pays for usually decides the cheapest option. Every ChatGPT plan is broken down in ChatGPT pricing in euros, and the Claude side in our Claude Code guide.

---

Where does your code go with Claude Code and Codex?

Both agents send the code they read to a US vendor for processing. Each therefore needs a data processing agreement and a short list of what the agent may open before any customer data goes near it. Anthropic is not on the EU-US Data Privacy Framework list, so a transfer assessment belongs in the file whichever agent you choose.

OpenAI offers European data residency on Enterprise and Edu, and its help centre lists the Codex app and command line as covered, while Codex on the web is not.¹³ Business customers can pick a storage region at checkout, but that choice does not cover processing, which runs elsewhere.¹³ The full picture is in ChatGPT data residency in Europe.

Anthropic runs no inference inside the EU itself. A company that must keep processing in Europe points Claude Code at Amazon Bedrock with EU cross-region inference instead. Called from an EU region, that route keeps each request inside six EU regions such as Frankfurt, Ireland and Stockholm.¹⁶ Cloud sessions do not run through Bedrock, so that route covers the local Claude Code only.⁴

---

How to choose between Claude Code and Codex in one week

• Pick one real task that takes a developer half a day, such as a bug with a test that reproduces it.
• Give both agents the same repository and the same prompt, each on its own branch, and write the house rules once in AGENTS.md, which both agents read.¹²
• Count three things: whether the tests pass, how many times a human had to step in and how long the review took.
• Check the licences you already pay for, because ChatGPT Business or Claude Team often settles the price.
• Sign the data processing agreement and list what the agent may see before any production code goes through it.

Expect both agents to get something wrong during the week, such as a test that passes for the wrong reason or a change that ignores a convention nobody wrote down. A developer still reviews every pull request before it merges, whichever tool you pick.

---

Summary: our verdict and next steps

For most of our customers we pick Claude Code, because it holds the top independent coding scores, connects to internal systems through MCP and shares one Team seat with the office tools. Codex is the better buy for a company that already pays for ChatGPT Business or wants a free tier to experiment. It also wins where EU data residency on Enterprise matters, or where large volumes of routine jobs make cost per task the deciding number.

Teams that pick Claude can have the licences and the rollout handled through satori-claude, at €490 once plus €14 per user a month, with the licence at €22 for Standard or €110 for Premium. A Claude workshop at €280 per user from eight users gets developers productive in two hours, and a Claude consultant can sit beside the team for the first weeks. Before any decision, a team can test the ground with a satori-launch at €560 per user from eight users. All satori. prices are fixed and exclude VAT.

Write one AGENTS.md file with your ten most important house rules, then run the same small bug fix through Claude Code and Codex on two branches. Within thirty minutes you will see which agent follows your rules without being reminded, and that tells you more than every table above.

---

Sources

Footnotes

1. Anthropic (2026). Claude Code overview. Claude Code Docs, checked on 9 November 2026. https://code.claude.com/docs/en/overview ↩ ↩² ↩³ ↩⁴

2. OpenAI (2026). ChatGPT and Codex documentation, checked on 9 November 2026. https://learn.chatgpt.com/docs ↩ ↩² ↩³

3. OpenAI (2026). Codex CLI. ChatGPT and Codex documentation, checked on 9 November 2026. https://learn.chatgpt.com/docs/codex/cli ↩

4. Anthropic (2026). Use Claude Code in the cloud. Claude Code Docs, checked on 9 November 2026. https://code.claude.com/docs/en/claude-code-on-the-web ↩ ↩²

5. OpenAI (2026). Agent approvals and security. ChatGPT and Codex documentation, checked on 9 November 2026. https://learn.chatgpt.com/docs/agent-approvals-security ↩

6. Anthropic (2026). Configure permissions. Claude Code Docs, checked on 9 November 2026. https://code.claude.com/docs/en/permissions ↩

7. Anthropic (2026). Models overview. Claude Platform Docs, checked on 9 November 2026. https://platform.claude.com/docs/en/models/overview ↩ ↩²

8. OpenAI (2026). Models. ChatGPT and Codex documentation, checked on 9 November 2026. https://learn.chatgpt.com/docs/models ↩ ↩²

9. OpenAI (2026). GPT-6.1 Sol. OpenAI API documentation, checked on 9 November 2026. https://developers.openai.com/api/docs/models/gpt-6.1-sol ↩

10. DeepLearning.AI (2026). Sol nearly eclipses Astra. The Batch. https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra ↩ ↩² ↩³ ↩⁴ ↩⁵

11. Anthropic (2026). Plans and Pricing, checked from an EU connection on 9 November 2026. https://claude.com/pricing ↩ ↩²

12. OpenAI (2026). Pricing. ChatGPT and Codex documentation, checked on 9 November 2026. https://learn.chatgpt.com/docs/pricing ↩ ↩² ↩³ ↩⁴

13. OpenAI Help Center (2026). Data residency and inference residency for ChatGPT, checked on 9 November 2026. https://help.openai.com/en/articles/9903489-data-residency-and-inference-residency-for-chatgpt ↩ ↩² ↩³

14. OpenAI Help Center (2026). ChatGPT Work and Codex, checked on 4 November 2026. https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex ↩

15. Anthropic (2026). Use Claude Code with your Pro or Max plan. Claude Help Center. https://support.claude.com/en/articles/11145838-use-claude-code-with-your-pro-or-max-plan ↩

16. Amazon Web Services (2026). Geographic cross-Region inference. Amazon Bedrock User Guide. https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html ↩