Aider

Almost every comparison of Aider against Cursor or Copilot starts in the wrong place, because Aider is not a competing product in the commercial sense. It is Apache 2.0 licensed software published on PyPI as aider-chat. There is no free tier because there is no paid tier. Aider AI LLC exists as the company behind the project, but it does not sell a hosted version or a pro plan.

What you pay for is inference. You supply an API key — Anthropic, OpenAI, Google, DeepSeek, xAI, Groq, Cohere, OpenRouter, Azure, Amazon Bedrock, Vertex AI, GitHub Copilot's endpoint, or anything OpenAI-compatible — and the provider bills you directly. Or you point Aider at Ollama or LM Studio, and the cost is zero plus whatever your GPU draws from the wall.

For a heavy daily user, this usually lands above a $20 subscription rather than below it, particularly with a reasoning model doing the editing. For occasional use it is dramatically cheaper, and for anyone who cannot send source code to a third-party vendor's servers, the bring-your-own-key architecture with a self-hosted or in-VPC model is the entire selling point.

Aider's most-copied idea is its repository map. Rather than embedding the codebase or dumping files into context, Aider parses the repo and builds a compact index of the important classes and functions with their signatures, showing the critical lines that define each symbol.

The clever part is selection. The full map is too large for most repos, so Aider treats source files as nodes in a dependency graph, runs a ranking algorithm over it, and sends only the highest-value portions that fit the token budget — 1k tokens by default via --map-tokens, expanded dynamically when no files have been added and the model needs a broader view of the codebase. The effect is that the model can see how the API of a module it is not editing actually works, and can ask for specific files when the map is not enough.

This is also why Aider's documentation actively argues against adding every file to the chat. The FAQ is explicit that dumping irrelevant files distracts the model, degrades edit quality and inflates your bill. You add the files that need to change; the map covers the rest.

/code is the default: describe a change and Aider edits the files in the chat. /ask discusses the code and answers questions without ever touching a file. /help answers questions about Aider itself. /architect is the interesting one.

In architect mode Aider makes two model calls. The main model acts as architect and proposes how to solve the problem; a second editor model turns that proposal into concrete file edits. You pick both — --model and --editor-model — which means you can put an expensive reasoning model on the thinking and a cheap, reliable model on the mechanical diff generation. Aider's documentation notes this was originally motivated by reasoning models that were strong at planning but poor at emitting precise edits, and that it can help even when both roles use the same model, simply by giving it two passes. The cost is two requests instead of one.

The documented alternative is the ask/code loop: use /ask to agree on an approach, then switch to /code and say "go ahead". Same effect, one model, and cheaper. Either way, the discipline of separating "decide what to do" from "write the diff" is the habit that makes Aider produce good changes.

Aider does not show you a diff pane because git already is one. Every edit it makes is committed with a message generated by the weak model from the diff and chat history, following Conventional Commits by default. If the files were already dirty, Aider commits your work first with its own message, so your changes and the AI's never end up in the same commit.

Attribution is built in: commits Aider authored get "(aider)" appended to the git author and committer metadata, commits it merely made from dirty files get it on the committer only, and flags exist to disable that, prefix messages with "aider: ", or add a Co-authored-by trailer instead. In-chat, /diff shows changes since your last message, /undo discards the last change outright, /commit commits dirty files, and /git runs raw git commands.

The result is that reviewing AI work becomes ordinary code review. You can bisect it, blame it, cherry-pick it, revert one commit out of five. No other tool in this category has made that quite so central.

Aider publishes two public benchmarks, and they became a de facto industry reference well beyond the tool's own user base. The older code editing leaderboard asks models to complete 133 Python Exercism exercises and apply every change to the source file without human intervention. It has been superseded by the polyglot leaderboard: 225 harder Exercism problems spanning C++, Go, Java, JavaScript, Python and Rust.

What makes those boards unusually honest is what they record alongside the pass rate. Each entry publishes the percentage of cases where the model produced a well-formed edit, the total API cost of the run, the exact command used, the edit format, the number of malformed responses, and seconds per case. The top polyglot entry as published is GPT-5 at high reasoning effort with 88.0 percent correct, 91.6 percent well-formed edits and $29.08 of API spend; o3-pro at high effort scores 84.9 percent with a 97.8 percent well-formed rate but costs $146.32.

The caveat is currency. The most recent run on the board is dated October 2025. As a historical record of how edit reliability and cost trade off, it remains valuable; as a guide to which model to point Aider at today, it has aged out.

Cursor and GitHub Copilot are editor-native. They give you inline completion, a diff review UI and multi-file context inside the environment you already work in, and they are what most developers should try first if they want AI assistance to feel invisible. Aider offers none of that, though it does have a --watch-files mode that picks up AI comments you leave in your editor.

Claude Code and Cline are agentic in a way Aider is not. They run shell commands, explore the filesystem, execute tests and iterate over long tasks with minimal supervision. Aider can lint and test automatically when you enable those flags, and can fix what the linter or test suite reports, but its default posture is a tight human-in-the-loop edit cycle rather than an autonomous agent.

Where Aider still wins outright: it costs nothing, it works with every model including local ones, its context strategy is genuinely economical, and its git integration is the best in the category. Where it loses: nothing about it is moving quickly any more. Commits still land in the repository — the most recent as of this writing were in May 2026, adding Anthropic model names and tree-sitter language support — but a project whose flagship benchmark has not been refreshed in nearly a year is not the one that will tell you which model to use next quarter.

Try it if you like working in a terminal, care about auditable commits, or need everything to stay on your own hardware. Look elsewhere if you want the tool to keep up with the models on its own.