If you still think of GitHub Copilot as the tool that completes your code as you type, you're looking at a version of the product that no longer exists. That's not a criticism — the change happened faster than most people's mental models about it. This article is an attempt to give you an accurate picture of what Copilot actually is today, what the real research says about its impact, and where the product is heading.
How it works under the hood
Every time Copilot generates a suggestion, it builds a prompt from all the context it can gather: the code around the cursor, other open tabs, the repository URL, custom instruction files, and — if you've configured them — indexed repository content or data from MCP servers. That prompt travels over TLS to GitHub's Copilot proxy, which handles authentication, content filtering, public code match checks, and rate limiting. It then routes to whichever model you've selected.
Inline suggestions use a Fill-in-the-Middle (FIM) paradigm, meaning the model sees both the code before and after the cursor, not just a prefix. GitHub ran A/B tests and found FIM raised accepted completions by around 10%. In 2024 they replaced the original completions backend with a custom-trained model that reduced latency by 35%, delivered a 12% higher acceptance rate, and tripled throughput.
For chat and agent workloads there's a model selector. As of April 2026, available families include OpenAI GPT-4.1, GPT-4o, and the GPT-5 family; Anthropic Claude Haiku 4.5, Sonnet 4/4.5/4.6, and Opus 4.5/4.6/4.7; Google Gemini 2.5 Pro and Gemini 3; xAI Grok Code Fast 1; and GitHub's own Raptor models. Older Claude 3.5/3.7 and Gemini 1.5/2.0 snapshots were deprecated in the second half of 2025.
The full feature landscape in 2026
Copilot expanded from one feature — inline suggestions — to something that looks more like a platform.
Next Edit Suggestions, available since April 2025 in VS Code, Xcode, and Eclipse, predicts where in the file you'll edit next, not just what comes after the cursor. It's a subtle difference but it changes how you navigate a codebase.
Copilot Edits, the multi-file editing mode, reached general availability in February 2025. It uses a dual-model architecture: one proposes the changes and a speculative decoding endpoint applies them quickly. You describe the task at a high level and the system touches whatever files are needed.
Agent mode is what changed the product's identity. It's available in VS Code, Visual Studio, JetBrains, Eclipse, and Xcode. In agent mode, Copilot chooses which files to modify, proposes terminal commands, executes them, reads the output, and iterates. It keeps going until the task is complete or it gets stuck. When GitHub announced it with Claude 3.7 Sonnet in April 2025, it published a 56% pass rate on SWE-bench Verified.
The cloud agent is the asynchronous version. You assign a GitHub issue to Copilot from the web, mobile, or CLI, and it runs inside a sandboxed GitHub Actions environment: it pushes commits to a draft PR, runs your tests, and requests review when it's done. It reached general availability in September 2025.
Copilot CLI reached general availability in February 2026. It's a separate install, available via npm, Homebrew, or WinGet, and brings a Plan mode, a fully autonomous Autopilot mode, parallel specialized agents (Explore, Task, Code Review, Plan), cross-session repository memory, hooks, plugins, and a built-in GitHub MCP server.
Copilot code review reached general availability in April 2025 and was redesigned at GitHub Universe 2025 to combine LLM reasoning with deterministic engines like ESLint and CodeQL.
Copilot Spaces, available since September 2025, are curated packages of files, issues, PRs, and documentation that act as baseline context for any Copilot surface.
For customization, you can configure a .github/copilot-instructions.md file at the repo level, instructions at the personal or organizational level, and since Universe 2025 an AGENTS.md file that defines custom agents with their own toolsets and per-project behavior.
Where Copilot runs
| Feature | Available in |
|---|---|
| Inline suggestions | VS Code, Visual Studio, JetBrains, Eclipse, Xcode, Vim/Neovim, Azure Data Studio |
| Chat | VS Code, Visual Studio, JetBrains, Eclipse, Xcode, GitHub.com, GitHub Mobile, Windows Terminal, Raycast |
| Agent mode | VS Code, Visual Studio, JetBrains, Eclipse, Xcode |
Vim/Neovim gets inline suggestions only, no chat. The CLI is cross-platform but not available on the free plan.
Plans and pricing, without the marketing
Copilot has five tiers. The central difference between them isn't so much the features as the premium request allocation — calls to frontier models, agent mode, and the cloud agent. Paid plans have unlimited completions and chat with included models. Overages on all paid plans cost $0.04 per additional premium request.
| Plan | Price | Completions | Premium requests |
|---|---|---|---|
| Free | Free | 2,000/month | 50/month |
| Student | Free (GitHub Education) | Unlimited | 300/month |
| Pro | $10/month | Unlimited | 300/month |
| Pro+ | $39/month | Unlimited | 1,500/month |
| Business | $19/user/month | Unlimited | 300/user |
| Enterprise | $39/user/month | Unlimited | 1,000/user |
One important practical note: Copilot is not available on GitHub Enterprise Server, only on GitHub Enterprise Cloud. That surprises many enterprise solutions architects.
What the real productivity data says
I want to be careful here because the numbers circulating online are frequently taken out of context.
The most cited study is Peng et al. (2022, arXiv:2302.06590). Ninety-five developers were randomly divided and asked to implement an HTTP server in JavaScript. The Copilot group finished in an average of 1 hour 11 minutes; the control group took 2 hours 41 minutes. That's a 55.8% speedup, statistically significant at P = 0.0017. Less experienced developers, older developers, and those with higher baseline workload benefited most. The task was specific and the sample was controlled, so this number describes a context, not all development work.
GitHub's randomized controlled trial with Accenture is the strongest enterprise evidence. Across roughly 450 Accenture developers, Copilot produced an 8.69% increase in pull requests per developer, a 15% increase in PR merge rate, and an 84% increase in successful builds. About 30% of Copilot's suggestions were accepted and 88% of accepted characters were retained.
A ZoomInfo field study (arXiv:2501.13282) covering over 400 developers found a 33% full-suggestion acceptance rate, a 20% line-of-code acceptance rate, and 72% developer satisfaction.
Numbers I'd avoid citing without a source: the famous "46% of code written by Copilot" and "15 million users" — those come from press releases, not controlled studies.
Where everything is heading
At GitHub Universe 2025, GitHub announced Agent HQ: a unified control plane that orchestrates agents from Anthropic, OpenAI, Google, Cognition, and xAI across GitHub, VS Code, CLI, and Mobile, all under a single Copilot subscription. The message was explicit: Copilot is positioning itself as the interface for all coding agents, not just GitHub's own.
The economic model is also changing. Each paid plan includes unlimited use of a base model with a monthly budget of premium requests for frontier model calls. As frontier models get cheaper, more of them will likely migrate to the base tier.
"Copilot is no longer a product — it's an orchestration layer. Completions, chat, edits, IDE agents, the CLI, and the cloud agent are points on a continuum that goes from 'suggest what comes next' to 'do this task and let me know when you're done.'"
— Carlos José Castro Galante
To wrap up
The most important thing to understand about Copilot in 2026 is that the underlying model changes constantly. What stays stable is the interface and, increasingly, the agents you define yourself. If you work with Azure or with development in general, getting familiar with how Agent HQ and the MCP ecosystem work today is more valuable than memorizing which model is available on which tier.
If you want to explore the Copilot learning resources Microsoft offers, Microsoft Learn has complete modules and learning paths covering everything from initial setup to agent mode and responsible AI.