Coding Assistants

Coding assistants, from the plan you pay for

Tools that write code with a model come in four shapes: an agent that works in your terminal, an agent built into a chat subscription, an extension for the editor you already use, and an editor of its own. They are priced differently and limit usage in different ways, so the price on the box rarely predicts your bill. The guides in this section compare the published plans of Claude Code, Codex, GitHub Copilot, Cursor, Google and Mistral, explain how to get started with the one most developers ask about, and mark the points where a local model is enough. We compare what the vendors publish and say what we have not measured: we have not run a controlled test of code quality, so we do not name a winner.

Where to start

1. Best coding assistant is the overview. It lists seven options by price and use: Claude Code with Claude Pro ($20 a month, or €18 before tax from Spain), Codex with ChatGPT Plus ($20, €23), GitHub Copilot (a free plan, then $10 for Pro), Cursor (a free plan, then from $8 a month), Google's Pro plan with Antigravity and Mistral Pro, plus a local model with Ollama when your code cannot leave your machine.
2. Claude Code vs Cursor vs Codex vs Copilot goes deeper on the four names most people compare. It shows what each plan includes, how each one limits usage (shared chat limits, requests, credits, messages) and why the shape of the tool matters as much as the price: Claude Code is an agent that runs in the terminal, IDEs, desktop app and web, while Cursor is a code editor built around an agent.
3. How to use Claude Code is a practical walkthrough: the one-command install on macOS, Linux, WSL and Windows, your first session in a project folder, the commands you will use every day and how a CLAUDE.md file gives the agent your project's rules before it starts.

• What is vibe coding? goes back to Andrej Karpathy's original post, explains where vibe coding ends and model-assisted development begins, and lists when it is safe and how to keep secrets, data and bills under control.
• AGENTS.md explained shows how to write the instructions file that Codex, Cursor, GitHub Copilot and Claude Code read, which file wins when there are several and how long it should be.
• Code Review Tools: Copilot, Claude Code and Codex compares GitHub Copilot code review, Claude Code /code-review and Codex: how to start each one, what it costs and how to get useful comments.

How to choose

Start with how you like to work, not with a leaderboard. If you live in the terminal and want an agent that reads a repository, edits files and runs commands, an agent such as Claude Code fits. If you already pay for a chat subscription, check first what it includes: Codex is part of ChatGPT, and Claude Code is part of the paid Claude plans, so the extra cost may be zero. If you want suggestions inside the editor you already use, GitHub Copilot has a free tier with a monthly limit of completions. If you want an editor designed around an agent, Cursor has a free plan with limited agent requests. Whatever you choose, test it for a week on your own repository, because the vendors' own documentation says usage depends on the size and complexity of the task.

What the price does not tell you

A plan price is a ceiling on what you pay, not a promise of what you get. Subscriptions share a usage allowance with chat, or count requests, or give a monthly credit, and a heavy agent session can spend an allowance quickly. Pay-as-you-go use through an API is billed per token, and agents repeat long prompts at every step, which is why prompt caching and our API price comparison matter if you build your own tooling. Plans change every few months, so each guide shows the date we read the vendor's page and a list of what may change.

Privacy and safety

Before you open a confidential project in a cloud tool, read the vendor's data settings. Individual plans and business plans often differ on whether your content may be used for training, as our guide on whether ChatGPT is safe shows for OpenAI. If your code cannot leave your computer, running a model locally is the alternative, and you can check whether the model fits your graphics card with our VRAM calculator. Agents also deserve care in general: they read files and run commands, so keep them on a Git branch, review every change before you commit it and give them only the permissions the task needs. The risks of giving an agent too much power are covered in what an agent is.

What to read next

If you use a coding agent with your own tools and data, read what MCP is. If your prompts are not getting the result you expect, our prompt engineering basics apply to code as much as to prose.