Claude Code vs. Codex: Which One Fits Your Team?

Your development team wants licences for an AI coding assistant, and two names are on the table: Claude Code from Anthropic and Codex from OpenAI. The question sounds like an IT decision you can delegate downwards.

Research it and you land on a results page full of developer blogs. They cover benchmarks, token consumption and terminal behaviour. Not one of them answers the three questions a managing director actually has: what does this cost, what are we risking, and how will we know afterwards whether it paid off.

Four in ten companies in Germany already use artificial intelligence. According to the Bitkom survey from March 2026 of 604 companies with 20 or more employees, the figure is 41 percent, with a further 48 percent planning a rollout or discussing one. For many of these companies, coding assistants are the first case in which AI does not merely write text but intervenes in a production system. That is precisely why this decision deserves a method.

What Claude Code and Codex actually are

Both tools belong to the second generation of AI coding aids. The first generation completed code in the editor, line by line, and the developer accepted or discarded the suggestion. Claude Code and Codex work differently: they read an entire project directory, change several files in one pass, run tests and, on request, produce a finished commit.

The difference is not a technical detail. It shifts the risk. A suggestion in the editor is harmless because a human checks it line by line. A tool that independently touches twenty files and reports the tests as green is, in practice, less often checked line by line. Comparing these two tools therefore means comparing two systems that act inside your own repository, not two autocomplete features. Here too, the jump from answering to acting is the point at which rules become necessary. We have described the difference between answering and acting AI systems in detail elsewhere.

What the benchmarks say, and what they do not

The public comparisons converge on a recurring picture: Claude Code is regarded as stronger in long sessions across many files, Codex as stronger on terminal and shell tasks and on delegating work to the background. That picture, however, comes from the experience reports of individual developers and trade blogs, not from controlled studies, and both vendors change their tools on a monthly cycle. As a working hypothesis for your own test, that is useful. As a purchasing decision, it is not.

How far experience reports can diverge from measurement is shown by a study from the research institute METR. Sixteen experienced open-source developers worked through 246 real tasks in their own projects, alternating with and without AI tools. Beforehand they expected a speed-up of 24 percent. They were measured 19 percent slower. And even after finishing the tasks, they still believed they had been 20 percent faster.

This study is not a verdict on all AI tools. It looked at experienced developers in projects they knew inside out, with an early-2025 generation of tooling. In an unfamiliar codebase or on routine work, the results may differ. What the study dismantles is a different assumption: that the people involved can reliably judge whether a tool makes them faster. They cannot. Anyone who wants to know the effect has to measure it.

The costs: near identical

The price comparison is quickly done, because there is barely anything to compare.

A standard seat on the Claude Team plan costs 20 US dollars per user per month billed annually, 25 US dollars billed monthly, and includes Claude Code. A standard seat on ChatGPT Business costs 20 US dollars per user per month billed annually, 25 US dollars billed monthly, with a minimum of two seats, and includes baseline access to Codex.

For a team of ten developers, both offers come to roughly 2,400 US dollars a year. At that order of magnitude the licence price is not a decision criterion but a rounding difference against the payroll cost of the same team.

The interesting line item sits above baseline access. Both vendors charge more once a team works intensively: Anthropic through premium seats at 100 US dollars per user per month billed annually, OpenAI through additional usage-based Codex seats. This line item is regularly missing from budgets. If you are planning one, ask the vendor for it up front rather than extrapolating from the base price.

The risk the comparison articles skip

Germany's Federal Office for Information Security (BSI), together with the French cybersecurity authority ANSSI, has published recommendations on the use of AI coding assistants. The two authorities name two risk areas.

The first is the confidentiality of the input. Whatever an assistant processes leaves the company: source code, configuration files and, in the bad case, credentials or personal data from test datasets. The second is the security of the generated code. Generated code can contain vulnerabilities, and it arrives in a form that looks competent in review. Neither is a theoretical objection. These are the points at which an unprepared rollout is noticed later.

On top of that comes an obligation many companies do not yet have on their list, and it changed a few weeks ago. Chapter I of the EU AI Act has applied since 2 February 2025 under Article 113(a), and it contains Article 4 on AI literacy. Since 27 July 2026 that article applies in a new version, amended by the Digital Omnibus Regulation on AI. Providers and deployers must now take measures to support the development of AI literacy among their staff. They expressly do not have to guarantee any particular level, which the law now says in so many words. The duty is therefore weaker than many training offers claim. It has not disappeared: an assistant handed out with no briefing at all is hard to defend as a supporting measure.

In practice that means three things before the rollout: a written rule on which data must not go into the tool, a data processing agreement where personal data is involved, and a short internal note on the briefing given. Article 4 does not require proof, but the note shows which measures you took. What such a rule looks like is set out in our guide to an AI policy for companies.

Three questions instead of a winner

Anyone deciding between Claude Code and Codex gets further with three questions than with any comparison test.

First question: what is allowed to leave your code base? Answer this first, because the answer can rule out both tools or restrict them to particular projects. For client work under a non-disclosure agreement, or for code that processes health or financial data, this is a legal question and not a technical one.

Second question: what work should the tool take on? A rebuild across many files in a legacy code base is a different job from a script that runs in the pipeline. Take the three task types your team actually handles most often and test both tools against them. Not against a demo.

Third question: what will you measure? Fix the metric in writing before the pilot. The useful figures are the ones already sitting in your systems: the time from ticket to merged state, the share of changes that come back in review, the number of defects noticed after deployment. The useless question is whether the team feels faster. The METR study shows why.

Hype or leverage?

Leverage, with one condition.

The tools are real, they are conspicuously cheap for the benefit they can deliver, and the two vendors barely differ on price. The hype is not in the tool but in the assumption that the choice between Claude Code and Codex is the important decision. It is not. Both tools will look substantially different in twelve months, and whoever buys one licence today can switch next year.

The important decision is whether you measure and govern the rollout, or whether you hand out licences and hope. The first costs two weeks of preparation. The second costs you the answer to the question of whether the rollout paid off.

The first step

Take one team and four weeks, starting Monday. Fix three things in writing beforehand: the three task types you will test against, the one metric from your own systems that shows the benefit, and the one page listing the data that must not go into the tool. Then give each tool two weeks. After four weeks you have a number instead of an opinion, and the decision makes itself.

If you would rather not set up that pilot alone: in the AI Potential Check we define the metric, the task types and the data rule with you, and support the evaluation. Get in touch.