AI for developers: coding agents, prices and what happens to your code

What coding agents are and who uses them

A coding agent is an AI program that does more than suggest code: it reads and changes files itself, runs commands and tests the result. It works in a repository, meaning the project folder with all the code and its version history. Many of these tools run in the terminal, the window where you type commands as text; others live in a code editor or a desktop app.

The best-known offerings are Claude Code from Anthropic, Codex from OpenAI, the Cursor editor, GitHub Copilot and Google's Antigravity CLI. The latter has replaced Gemini CLI for personal users: according to Google, Gemini CLI has stopped serving requests for free users and for Google AI Pro and Ultra subscribers since 18 June 2026. It remains usable with a Gemini Code Assist Standard or Enterprise licence and with paid API keys.

In the Stack Overflow Developer Survey 2026, 66 per cent of respondents said they use coding assistants or coding agents at work. When asked about coding agents, Claude Code led with 66 per cent and GitHub Copilot with 59 per cent. If you cannot program at all and want to build a first small website, there is a separate article on coding agents for non-programmers.

What the tools cost (as of October 2026)

Many developer tools only state their prices in US dollars. We give them as the vendors state them and do not convert them.

According to Anthropic, Claude Code is included in all paid Claude plans but not in the free version. Pro costs 17 dollars a month billed annually or 20 dollars billed monthly, and Max starts at 100 dollars, each excluding tax. According to OpenAI, Codex is included in all ChatGPT plans, with limits in Free and Go. According to OpenAI, Codex Cloud, meaning tasks that run in OpenAI's cloud, is not part of Free and Go. In Germany, ChatGPT Go costs 8 euros, Plus 23 euros and Pro from 103 euros a month, according to OpenAI.

Cursor has a free Hobby plan with limited agent requests; Pro costs 20 dollars a month plus tax, Teams 40 dollars per person. GitHub Copilot is available free of charge with limited use, Pro costs 10 dollars, Pro+ 39 dollars and Max 100 dollars a month. For companies, GitHub lists 19 dollars per seat for Business and 39 dollars for Enterprise.

According to Google, Google Antigravity is free for individuals, with weekly usage limits. More quota comes with Google AI Pro, which costs 21.99 euros a month in Germany, or with Ultra. All vendors also limit usage through quotas. Depending on the vendor and plan, they apply per few hours, per week or per month; GitHub Copilot Free, for example, allows 2,000 code completions a month according to GitHub.

What happens to your code

The key difference is between personal and business plans. On Claude Free, Pro and Max, according to Anthropic, you decide whether your data, including from Claude Code, is used to train future models. If you allow it, Anthropic keeps it for up to five years, otherwise for 30 days. On Team, Enterprise and the API, Anthropic says it does not train on code or prompts.

On ChatGPT Plus and Pro, according to OpenAI, conversations may be used for training unless you switch this off in the data controls; the setting also applies to Codex. On Business, Enterprise and the API, OpenAI does not use inputs by default. OpenAI's pricing page does not list data residency in the EU for the personal plans.

Since 24 April 2026, GitHub has used inputs, outputs and code snippets from Copilot Free, Pro, Pro+ and Max for training unless you object in the Copilot settings. heise reported on the switch to an opt-out approach. According to its documentation, GitHub does not use data from Copilot Business and Enterprise for training.

According to the vendor, Cursor does not use your code for training when Privacy Mode is switched on; without it, Cursor may use code data for improvement and training. For teams, Privacy Mode is on by default. For Antigravity for individuals, Google's terms say it uses your interactions to improve its products and AI and that staff may review them; this can be changed in the settings.

Security: protect secrets, check the code

Germany's BSI and its French partner agency ANSSI have published joint recommendations for AI programming assistants. They count the confidentiality of inputs and security flaws in generated code among the risks.

This leads to concrete rules. Credentials such as API keys, the passwords programs use to access services, do not belong in inputs to the AI. Generated code should be checked and understood by developers, particularly with regard to hallucinations and security flaws. A known risk is invented package names: attackers publish malware under exactly those names. The agencies also warn of indirect prompt injections, meaning hidden instructions in files or web pages that manipulate the AI's behaviour.

For companies, the BSI and ANSSI recommend not allowing private accounts for developing company software and instead offering securely configured business accounts with clear rules.

Developers themselves are cautious too: in the Stack Overflow Survey 2026, 48 per cent said they trust AI when they can easily verify the result. Our assessment: at the start, have every change shown to you, check it yourself and only then approve it. Also give the agent access only to the project it should work on.

Which tool for whom

Our assessment, not a test verdict: if you already have a ChatGPT or Claude subscription, you can try Codex or Claude Code at no extra cost. If you prefer working in an editor with suggestions as you type, Cursor or Copilot is the right choice. Copilot Free and Antigravity for individuals are suitable for free testing but give considerably less quota.

For code belonging to clients or employers, business plans are the safer choice, because according to the vendors they do not train by default. On personal plans, always check the training setting yourself before entering other people's code.