If you still think of GitHub Copilot as "the thing that autocompletes your code," you're about two years behind. That's not a criticism - the product has changed faster than most people's mental models of it. This post is an attempt to give you an accurate picture of what Copilot actually is right now, what the research says about its impact, and where the real limits are.
What it does under the hood
Every time Copilot generates a suggestion, it builds a prompt from whatever context it can gather: the code around your cursor, other open tabs, your repo's URL, any custom instruction files you've set up, and - if you've configured them - indexed repository content or attached data from MCP servers. That prompt goes over TLS to GitHub's Copilot proxy, which handles authentication, content filtering, public-code-match checks, and rate limiting. Then it routes to whatever model you've selected.
Inline completions use a Fill-in-the-Middle (FIM) approach, meaning the model sees both the code before and after the cursor rather than just a prefix. GitHub ran A/B tests on this and found it lifted accepted completions by around 10%. In 2024 they also swapped out the original completion backend for a custom-trained model that reduced latency by 35%, delivered 12% higher acceptance rates, and tripled throughput.
The feature surface in 2026
Copilot has expanded from one feature (inline completions) 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're going to edit next, not just what comes after the cursor. It's a subtle difference but it changes how you move through a codebase.
Copilot Edits / multi-file edit mode reached GA in February 2025. It uses a dual-model architecture: one model proposes the changes, a speculative-decoding endpoint applies them fast. You describe what you want at the level of a task, and it touches as many files as 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 picks the files to touch, proposes terminal commands, runs them, reads the output, and iterates. It keeps going until the task is done or it gets stuck. When GitHub announced it with Claude 3.7 Sonnet in April 2025, it posted a 56% pass rate on SWE-bench Verified.
The cloud agent (launched GA in September 2025) is the async version. You assign a GitHub issue to Copilot from the web or CLI, and it runs inside a sandboxed GitHub Actions environment, pushes commits to a draft PR, runs your tests, and requests your review when done. You don't have to be at your desk.
The Copilot CLI reached GA in February 2026. It's a separate install (npm, Homebrew, or WinGet) that brings a Plan mode, a fully autonomous Autopilot mode, parallel specialized sub-agents (Explore, Task, Code Review, Plan), repository memory across sessions, hooks, plugins, and a built-in GitHub MCP server.
Copilot code review reached GA in April 2025 and was rearchitected at GitHub Universe 2025 to combine LLM reasoning with deterministic engines like ESLint and CodeQL. In December 2025 it was extended so that PRs from unlicensed contributors in an org can still be reviewed, billed to the org.
Copilot Spaces (GA September 2025) are curated bundles of files, issues, PRs, and docs that act as grounding context for any Copilot surface.
On customization: you can set a .github/copilot-instructions.md at the repo level, personal or org-level instructions, and since Universe 2025 an AGENTS.md file that defines custom agents with their own tool sets and behavior per project. MCP has become the primary extension mechanism - servers get invoked automatically based on intent rather than requiring explicit calls.
Where it runs
Inline completions are supported in VS Code, Visual Studio, JetBrains IDEs, Eclipse, Xcode, Vim/Neovim, and Azure Data Studio. Chat runs in VS Code, Visual Studio, JetBrains, Eclipse, Xcode, GitHub.com, GitHub Mobile, Windows Terminal, and Raycast. Agent mode is in VS Code, Visual Studio, JetBrains, Eclipse, and Xcode. Vim/Neovim gets completions only - no chat. The CLI is cross-platform but not available on the Free tier.
Plans and pricing, without the marketing
There are five tiers. The Free plan launched in December 2024 with 2,000 completions and 50 premium requests per month. The Student plan is free for verified GitHub Education users and gives unlimited completions with 300 premium requests. Pro is $10/month (or $100/year) with unlimited completions and 300 premium requests - it includes agent mode, the cloud agent, the CLI, and MCP. Pro+ is $39/month with 1,500 premium requests and access to every available model, including preview models as they ship. Business is $19/user/month (300 premium requests per user, admin controls, audit logs, IP indemnity). Enterprise is $39/user/month (1,000 premium requests per user, organization-codebase indexing, a fine-tuned private completion model, and Bing-grounded web search in GitHub.com chat).
Overages on every paid plan cost $0.04 per additional premium request. Base-model usage - inline completions, chat with the included models - doesn't count against the premium budget.
One practical note: Copilot is not available on GitHub Enterprise Server, only on GitHub Enterprise Cloud. That surprises a lot of enterprise architects.
What the productivity data actually says
I want to be careful here because the numbers that circulate online are often decontextualized.
The most cited study is Peng et al. (2022, arXiv:2302.06590). Ninety-five developers on Upwork were randomly split and asked to implement an HTTP server in JavaScript. The Copilot group finished in 1 hour 11 minutes on average; the control group took 2 hours 41 minutes. That's a 55.8% speedup, statistically significant (P=0.0017). Less experienced developers, older developers, and those with higher baseline workloads benefited most. The task was narrow and the sample was controlled, so this number describes one context, not all development work.
The GitHub × Accenture randomized controlled trial 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 suggestions were accepted, and 88% of accepted characters were retained. Accenture has since rolled Copilot out to more than 12,000 developers.
A ZoomInfo field study (arXiv:2501.13282) covering 400+ developers through a four-phase rollout found a 33% full-suggestion acceptance rate, a 20% line-of-code acceptance rate, and 72% developer satisfaction.
The numbers I'd avoid citing without sourcing are the "46% of code written by Copilot" and "15 million users" figures - they come from press announcements rather than controlled studies. The Forrester ROI figures are real but behind a paywall; if you want to cite them, get the original study.
The direction things are going
At GitHub Universe 2025, GitHub announced Agent HQ, a control plane that orchestrates agents from Anthropic, OpenAI, Google, Cognition, and xAI across GitHub, VS Code, CLI, and Mobile under a single Copilot subscription. The framing was explicit: Copilot is positioning itself as the interface for all coding agents, not just the home for GitHub's own.
The economic model is also shifting. Every paid tier includes unlimited use of a base model with a monthly premium-request budget for frontier calls. As frontier models become cheaper, more of them will probably move into the base tier. For now, the budget disciplines how much you use the most powerful models per month.
If there's one sentence that captures where Copilot is in 2026: it's not a product anymore, it's an orchestration layer. Completions, chat, edits, in-IDE agents, the CLI, and the cloud agent are points on a continuum from "suggest what comes next" to "go do this task and tell me when you're done." The underlying model changes constantly. What stays stable is the interface - and increasingly, the agents you define yourself.
If you want to dive deeper into Copilot's learning resources, Microsoft Learn has a full set of modules and learning paths covering everything from setup to agent mode and responsible AI:
👉 https://learn.microsoft.com/copilot?wt.mc_id=studentamb_510659
Carlos José Castro Galante is a Full Stack Developer and Azure AI Engineer certified by Microsoft (AI-102, AI-900, AZ-900) and ITBA. Available for freelance projects from Argentina.
Top comments (0)