Overview
GitHub Copilot started as a Codex-powered autocomplete and has grown into a full IDE assistant with chat, edits, workspace, and Copilot Workspace / Agent modes. Because it's bundled into GitHub billing, it's the AI most professional developers actually have available at work. Model choice now spans GPT-5, Claude Sonnet 4/Opus 4, Gemini 2.5 Pro, and o-series reasoning models.
Capabilities
- Inline completions in every major editor (VS Code, JetBrains, Vim, Xcode, Visual Studio).
- Chat with codebase context — Copilot Workspace pulls repo files automatically.
- Multi-file edits and terminal command generation.
- Copilot for Pull Requests: summaries, review, and test suggestions.
- Model routing across GPT, Claude, Gemini families.
Where it shines
- Best autocomplete UX in-editor. Latency and diff quality are hard to beat.
- Enterprise-grade compliance: SOC2, no code training on business/enterprise tiers.
- Deepest integration with GitHub — PR reviews, issue triage, workspace context all just work.
Where it struggles
- Agent modes (Workspace, multi-file edits) still trail Cursor's Cmd-K and Composer.
- Chat context can be smaller than the underlying model's max — the plumbing throttles it.
- Custom instructions are per-repo, not per-project or per-user, which is awkward for polyglot teams.
Best for
- Teams already committed to GitHub for source control.
- Enterprises that need SOC2 / no-training guarantees out of the box.
- Developers who spend most of their day in an editor autocompletion-first workflow.
Pricing notes
Business ~$19/user/mo, Enterprise ~$39/user/mo. Compared to per-token API costs for the same models, the flat-rate math wins for most teams once usage exceeds a few hours a week.
Tutorials & patterns that work
Configure repo custom instructions
Add .github/copilot-instructions.md with project conventions: file layout, naming, test pattern, forbidden libraries. Copilot chat and edits use it automatically. This is the single biggest quality lever most teams ignore.
Use Copilot for PR reviews without ceding judgment
Enable Copilot review on PRs, but treat its comments as a first-pass linter, not an approver. It catches obvious bugs and missing tests; humans still own architectural judgment.
Model selection inside chat
For refactors involving many files → switch to Claude Sonnet. For math/algorithms → o-series. For quick answers → GPT-4o-mini. Copilot's model picker is per-conversation, not global.