GitHub Copilot

The choice rests on two things: where your code lives and where you work from.

Who it is for. For those whose projects are already on GitHub: suggestions, chat, and a cloud agent are enabled in a couple of clicks and count as one subscription. Even more so for teams: at companies with 10,000+ people, Copilot has 56% usage, more than any competitor (Pragmatic Engineer survey, 02.2026).

Who it is not for. Those who need full freedom over the model. Copilot accepts your own key only in editors, while any model anywhere is available in the open-source Cline and OpenCode.

Starting from scratch? Signing up here is easier than with many others: a free plan and suggestions right in the editor. But to assign tasks to the agent and review the result, start with the vibecoding page — the door is below.

It works — both entry points are in section 3. Choosing between it and Cursor: the decision is in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

One tool, four hands: from suggestions at your fingertips to an agent that works while your laptop is closed.

Suggestions. Copilot completes code as you type and suggests the entire next edit. This is the only part without a counter: completions are unlimited on paid plans and spend no credits.

Chat and agent mode. In the editor, ask in plain language: explain, fix, add a test. In agent mode, Copilot finds the needed files itself, runs commands, and shows a finished set of edits with buttons to accept or revert.

Cloud agent. You assign an issue to @copilot, and the agent works in GitHub's cloud: it edits the code in a separate environment, runs the tests and opens a pull request for your review. Tasks run in parallel, and a single session tops out at 59 minutes (verified Sep 26, 2026).

Terminal. Copilot CLI installs with one command from section 3 and is included in every plan, including the free one. Alongside its own modes, Copilot also lets you use third-party agents in GitHub: Claude Code and Codex connect through the same subscription.

There are two doors: the editor for daily work and github.com for agent tasks.

Editor. Copilot is already built into VS Code: sign in with a GitHub account, and the free plan activates without a card. Plugins are available for Visual Studio, JetBrains, Eclipse, and Xcode.

Site and phone. On github.com open chat or assign an issue to @copilot: the cloud agent needs no installation at all. GitHub Mobile does the same on your phone.

How to install the terminal version

Step 1. Install the agent with one command:

npm install -g @github/copilot

Step 2. Run the copilot command and sign in with your GitHub account. The CLI works on macOS, Linux, and Windows and is included in every plan, including Free.

Install command verified Sep 26, 2026 against docs.github.com (Install Copilot CLI)

An agent benchmark is always a benchmark of an “agent + model” pair. Copilot has dozens of models in its menu, and there is no official pair in benchmarks: it has no single score.

Scale. More than 20 million people over its lifetime (words of Microsoft CEO Satya Nadella, July 2025). GitHub counted 150 million developers on the platform back in December 2024, so roughly every seventh one has tried Copilot. 4.7 million pay for it, up 75% year over year (Microsoft report, January 2026).

Love and scale diverge. In the Pragmatic Engineer survey (906 respondents, February 2026), 9% named Copilot their “favorite tool,” versus 46% for Claude Code. But at companies with 10,000+ employees, it leads usage at 56%: they love other tools, but work in this one.

You pay for a plan, and credits are counted within it: since Jun 1, 2026, the whole lineup uses one currency, AI Credits, with one credit equal to one cent.

What spends credits. Chat, agents, code review, and terminal use, based on tokens from the selected model: a heavy model burns through the monthly bundle faster. Completions spend no credits at all.

What models come with the plan. Pro at $10 gives you workhorse models like Sonnet 5.5, while the flagships Claude Opus 5.5 and GPT-6 Astra unlock with Pro+ at $39. On Free, the model is picked automatically. The full list with retirement dates for older models is in the FAQ.

Free plan. Permanent and no card required: 2,000 completions per month plus limited chat with the agent. Enough to try it, tight for daily work.

We look at three things: where your code goes, what happens with training on it, and what is visible in the tool itself.

Code goes to the cloud. Suggestions and chat work through GitHub servers, while the cloud agent runs tasks in GitHub Actions. There is no lock-in: the agent's changes arrive as a regular pull request to your repository, so you do not need to retrieve anything separately.

Training on your data is enabled by default. Since Apr 24, 2026 on Free, Pro, Pro+ and Max plans, conversations with Copilot, including code snippets, are used to train GitHub models unless you turn this off in settings. Business and Enterprise plans are protected by contract: their data is not used for training.

The tool itself is closed. The core and models are not published, and cannot be self-hosted. Editor plugins and the source of the chat extension for VS Code are open. The copilot-cli repository is a distribution showcase with complaint intake, not an open core.

The "github copilot vs" search tail almost always leads to Cursor. Short answer: Copilot adds an agent where the code already lives, while Cursor builds an entire editor around the agent.

Entry price. Both have a free plan, while Copilot Pro is half the price: $10 versus $20. The credit math is the same: heavy models burn through the package faster.

Measurements. In the July version of the Artificial Analysis index the best Cursor CLI pair, with GPT-5.5, scored 62 points; in the September one there are no Cursor pairs at all. Copilot has no measured pair whatsoever - the ceiling is set by the model you pick from the menu. In the Pragmatic Engineer survey Cursor is liked more often (19% vs 9%), while at large workplaces Copilot is installed more often: 56% at companies with 10,000 employees and up.

Who else is nearby. For depth in the terminal people go to Claude Code (its best pair, with Opus 5.5, scores 66 AA points), for your own key and open source - to Cline and OpenCode.

The product is five years old, and it is the elder of the genre: it went from the world's first mass-market code suggestion to a lineup where the agent works in the cloud without a person at the keyboard.