The short answer
GitHub Copilot is GitHub's AI coding assistant for editors such as Visual Studio Code and JetBrains IDEs, the terminal, a desktop app and github.com. It suggests code, answers questions about your codebase and runs agents that edit, test and open pull requests. Since 1 June 2026 all plans bill by GitHub AI Credits, and since 28 October 2026 Copilot Fusion can pick models for each task. Copilot Free includes 2,000 completions a month.
| Made by | GitHub, United States |
|---|---|
| Free plan | Yes: Copilot Free, 2,000 completions a month, Auto model only, no cloud agent |
| Models | Copilot Fusion and Auto, plus GPT-6 and GPT-6.1, Claude Opus 5.5 and Sonnet 5.5, Gemini 4 Argon and others |
| Best at | Inline suggestions, agent tasks from issue to pull request, code review |
| Shows sources | It points to files and lines in your code; it is not a web research tool |
| Team plans | Copilot Business and Copilot Enterprise, plus pay-for-credits access for GitHub Enterprise |
| Your data | Free, Pro, Pro+ and Max interactions can train models unless you opt out. Business and Enterprise are excluded. |
What GitHub Copilot is in late 2026
GitHub Copilot started as autocomplete for code. It is now a family of tools that share one account: inline suggestions in your editor, a chat that knows your codebase, an agent mode that edits several files and runs commands, a cloud agent that works on GitHub while you do something else, AI code review on pull requests, a terminal tool called Copilot CLI and a desktop app. It works in Visual Studio Code, Visual Studio, JetBrains IDEs, Xcode and Eclipse, on github.com and in GitHub Mobile.
Three changes this year shape how you should use it. First, on 1 June 2026 every plan moved to usage-based billing with GitHub AI Credits, priced by the tokens each model uses at its published API rate. Code completions and Next Edit suggestions are still included and do not use credits. Second, the GitHub Copilot app became generally available on 17 June 2026 for macOS, Windows and Linux, and since 7 July it is on every plan, including Copilot Free. Third, at GitHub Universe on 28 October 2026, GitHub made Copilot Fusion generally available, a router that picks which models handle each task.
Which models GitHub Copilot uses
Copilot is unusual because it is not tied to one AI company. GitHub's supported models list includes OpenAI's GPT-6 Sol, GPT-6 Luna, GPT-6 Astra, GPT-6.1 Sol and GPT-6.1 Terra, Anthropic's Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5, Google's Gemini 4 Argon and Gemini 3.8 Flash, xAI's Grok models, Moonshot AI's Kimi K3 and Microsoft's MAI-Code-1.1-Flash. GPT-6.1 Terra arrived on 26 October 2026 for Pro, Pro+, Max, Business and Enterprise, aimed at routine agent and terminal work at a lower cost than GPT-6.1 Sol. GPT-5.3-Codex is the long-term support model that Copilot falls back to when nothing else is available.
You do not have to choose. Auto picks a model for you, and it is the only option on Copilot Free and Copilot Student. Copilot Fusion goes further: it decides per task whether one model should do the job (Single), whether an efficient model should draft and a judge accept it or escalate (Cascade), or whether several models should critique a draft in parallel before it is revised (Critique). You choose an effort level. Fusion Standard is on every plan. Fusion High and Fusion Max need Pro+, Max, Business or Enterprise. There is no separate Fusion charge, you pay for each model it calls, and through 30 November 2026 paid plans get a 30% discount on model costs when using Fusion.
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Claude Opus 5.5 Max, at about half the average API cost per task. That figure comes from GitHub itself, so test it on your own work before you change how your team works.
What GitHub Copilot is best at
Suggestions while you type
Inline completions are still the feature most developers use all day, and they cost nothing in credits. Copilot Free includes 2,000 completions a month. Writing a clear comment or function name first gives it the context to suggest the right code.
Agent work inside GitHub
This is Copilot's real advantage over editor-only tools. Since 28 October 2026, the new Copilot in the cloud experience on github.com and GitHub Mobile, for paid plans, can research a repository, plan changes, edit files and run tests and linters in a cloud environment, then open a pull request for you to review. You can ask it to fix failed workflows or merge conflicts, and you can assign it code scanning alerts. With session continuity, a cloud agent session you start in the app can be followed in VS Code, the CLI or on the web.
Work across repositories and on a schedule
Copilot projects, in the Copilot app, run agent sessions across several repositories in parallel with one place to track them. Copilot automations run workflows in a repository manually, on an event or on a schedule, for jobs like issue triage and code review.
Code review on pull requests
Copilot code review comments on pull requests at three levels, Lite, Balanced and Max. From 28 September 2026 the Default effort uses Balanced rather than Lite, unless an admin picked Lite. Since 1 June 2026 code review also uses GitHub Actions minutes as well as AI credits, so budget for both.
Where GitHub Copilot falls short
- Costs are harder to predict. Chat and agent runs spend AI credits by token, and input, output and cached tokens all count, so long prompts, large context and long answers use more. Since June there is no automatic fallback to a cheaper model when credits run out, so a long agent session on an expensive model can use a month's allowance fast.
- The free plan is narrow. Copilot Free gets Auto model selection only, 2,000 completions a month, no cloud agent, no third-party agents and code review only through Review selection in VS Code.
- Agents have hard limits. GitHub's docs say the cloud agent works on one branch and opens one pull request per task, sessions on GitHub Actions stop at 59 minutes, and only repositories hosted on GitHub are supported.
- You still own the code. GitHub warns that Copilot can produce incorrect code, including code with security vulnerabilities. With Fusion, edits from a discarded draft are not undone automatically, so review every diff before you commit.
How to get better results from GitHub Copilot
- Pick the right surface for the job. Completions for the line you are writing, chat for questions, agent mode for a change across a few files, the cloud agent for a well-defined issue you would hand to a colleague.
- Write tasks like tickets. State the goal, the files or modules involved, the constraints ("do not change the public API") and how to verify it ("all tests in /billing must pass").
- Ask for a plan first. For anything bigger than a small fix, have Copilot list the steps and files it will touch, correct the plan, then let it write code.
- Start with Fusion Standard or Auto. Move to a named frontier model or a higher Fusion level only for the hard tasks. Compare credits used and how much you had to fix by hand.
- Keep the context tight. Close unrelated files and point to the exact functions. Bigger context costs more credits and rarely helps.
- Make tests the judge. Ask Copilot to write or update tests with every change and to run them. A passing test suite is a better check than a confident explanation.
Our GitHub Copilot lessons practise these habits, and Cursor vs GitHub Copilot covers the main rival.
Privacy settings: the training switch to check today
Since 24 April 2026, GitHub may use your interactions with Copilot to train and improve AI models if you are on Copilot Free, Pro, Pro+ or Max, unless you opt out. To turn it off, click your profile picture in the top right of any GitHub page, choose Copilot settings, open Allow GitHub to use my data for AI model training and select Disabled. If you do not see that setting, you are probably signed in with a Copilot Business or Enterprise licence, whose data is not used for training.
Two more details matter for teams. Under the unified Copilot experience that launched no earlier than 28 September 2026, chat data on github.com is kept for the life of the account rather than 28 days. And on Business and Enterprise, admins decide which models are enabled through model policies. Under the Default availability for released models policy, newly released generally available models are switched on automatically unless an admin disables that policy or the individual model. Pre-release models, open-weight models such as Kimi K3 and models outside GitHub's data retention agreement, such as Claude Fable 5, stay off until an admin enables them. If your company handles client code, use a Business or Enterprise seat rather than a personal plan.
Which GitHub Copilot plan suits whom
- Copilot Free: for trying it, hobby projects and students who are not yet verified. Completions, chat on github.com and the Copilot app, with Auto model only.
- Copilot Student: for verified students. Adds the cloud agent and code review, still with Auto model only.
- Copilot Pro: for individual developers who code most days. 1,500 AI credits a month (1,000 base plus 500 flex) and a choice of models.
- Copilot Pro+: for heavy agent users. 7,000 AI credits a month, access to premium models and Fusion High and Max.
- Copilot Max: for people who run agents all day. 20,000 AI credits a month and priority access to premium models.
- Copilot Business and Enterprise: for teams. 1,900 and 3,900 AI credits per user a month, admin policies, budgets and no training on your data. Since Universe, GitHub Enterprise customers can also use every Copilot feature without a separate Copilot subscription and pay only for the credits they use, through GitHub Sales for now.
Who it is not for: if you do not write or review code, Copilot is the wrong tool, and Microsoft Copilot in Office is a different product. Check GitHub's plans page for current prices, and see best AI for coding and compare for alternatives.
Prompts to copy
Explain an unfamiliar codebase
I am new to this repository. Explain how it is structured: the main entry points, how a request flows from [entry point] to the database, and the five files I should read first. Point to file paths for every claim.
You get: A map of the codebase with file paths to check
Plan before you change anything
I want to [describe change, e.g. add rate limiting to the public API]. Do not write code yet. List the files you would change, the steps in order, the risks, and the tests that should prove it works.
You get: A reviewable plan you can correct before any edits
Fix a bug with a test first
Users report [bug description]. Write a failing test that reproduces it, run it to confirm it fails, then fix the bug with the smallest change possible and run the full test suite. Summarise what caused it.
You get: A fix proven by a test, plus the root cause
Hand an issue to the cloud agent
Implement issue #[number]. Constraints: do not change the public API, follow the patterns in [folder], and keep the change under [N] files. Add or update tests, make sure linting passes, and open a pull request that explains each change.
You get: A pull request you can review like a colleague's
Review a pull request before a human does
Review the changes in this pull request for bugs, security issues, missing error handling and missing tests. Rank the findings by severity, quote the lines, and suggest a fix for each. Skip style nitpicks.
You get: A ranked list of real problems with suggested fixes
Write tests for untested code
Write unit tests for [file or function]. Cover the normal case, edge cases such as empty input and very large values, and the error paths. Use our existing test framework and naming style, then run them.
You get: A test file that matches your project's conventions
Refactor safely
Refactor [function or module] to [goal, e.g. remove the duplicated validation logic]. Behaviour must not change. Run the existing tests before and after, and list anything you were unsure about.
You get: Cleaner code with tests showing nothing broke
Turn a command-line task into a script (Copilot CLI)
Find every log file in [folder] older than 30 days, compress them into one archive named by today's date, and print how much space this saves. Show me the commands and explain each one before you run anything.
You get: Commands explained before they run