GitHub Copilot

What it actually is

Copilot works as four layers stacked on top of each other, each trading speed for autonomy. Code completion is the original feature — ghost-text suggestions as you type, unlimited on every plan including Free. Copilot Chat answers questions about your code conversationally, in the editor or on GitHub.com. Agent mode takes a plain-language task, reads across multiple files, makes the edits, runs terminal commands, and self-corrects when something breaks — generally available in VS Code and, since March 2026, at full feature parity in JetBrains too. The coding agent goes further still: assign it an actual GitHub issue and it works asynchronously in a sandboxed environment, opening a draft PR for review while you do something else.

None of this is locked to one model. You can route a task to GPT, Claude, or Gemini specifically, or let auto-selection choose for you. Pricing sits on top as six tiers — Free, Pro ($10), Pro+ ($39), Max ($100), Business ($19/seat), and Enterprise ($39/seat) — and as of June 1, 2026, GitHub replaced its old Premium Request Units with dollar-denominated AI Credits (1 credit = $0.01). Completions stay unmetered; chat, agent mode, the cloud agent, code review, and the CLI all draw from a monthly credit pool sized to your plan.

Where it’s genuinely good

The distribution advantage is real and easy to underrate. You don’t adopt a new editor to use Copilot — it’s an extension in the tool you already have open, which is a meaningfully lower-friction start than a VS Code fork asking you to migrate your whole setup. The agent stack has also matured fast: agentic code review shipped in March 2026 and can hand a flagged problem directly to the coding agent to generate a fix, which closes a loop that used to require a human to notice, describe, and delegate the fix manually. Model choice adds real flexibility too — routing a large refactor to a stronger reasoning model while leaving routine completions on a cheaper one is a genuine cost lever, not just a settings toggle nobody uses.

Free access through GitHub Education and the Global Maintainers Program is also worth knowing about specifically — it’s not a trial, it’s full Pro access at no cost for two large, identifiable groups, which is more generous than most competitors offer anyone.

Where it falls short

The vulnerability research is the part worth taking most seriously, precisely because it’s not a hot take — it’s measured. Academic analysis of real Copilot suggestions merged into actual GitHub projects has repeatedly found security weaknesses in something like a quarter to a third of accepted snippets, and separate research has shown Copilot will replicate a vulnerability already sitting in your codebase rather than flag it. The most pointed recent finding is also the most current: a 2026 study specifically testing the new agentic code review feature against known-vulnerable code found it often missed SQL injection, XSS, and insecure deserialization — the marquee categories that feature exists to catch. None of this means the tool is unsafe to use. It means the “AI reviewed it” step isn’t a substitute for a human security review, and treating it as one is exactly the failure mode the research is describing.

The billing shift is the other thing to plan around, not just note. Moving from Premium Request Units to AI Credits in June 2026 kept headline prices the same but changed what they buy — heavy agent-mode or cloud-agent usage with frontier models can burn through a monthly allowance faster than the old system implied, especially on Pro+ if you’re specifically paying for Claude or o3-class model access. And the Doe v. GitHub litigation, while narrowed substantially since 2024, hasn’t resolved the actual question at its center: whether Copilot’s output can constitute a derivative work under the open-source licenses of the code it was trained on. That’s still an open legal question, not settled history.

Which plan to actually get

Free is a genuinely reasonable way to try it — a generous completions allowance and real, if limited, chat and agent access, not a crippled demo. Move to Pro once you’re using it daily; at $10/month it’s priced to be a low-friction yes, and it covers the large majority of individual developers without hitting a wall. Pro+ only earns its $39 if you specifically want the heavier models — Claude Opus-class or o3-class — on a regular basis; check your first month’s credit usage before assuming you need it. Max, at $100, is built for people running agents for hours a day, not occasional heavy sessions — new sign-ups have also been intermittently paused in 2026, so confirm availability before budgeting around it. For teams, Business at $19/seat covers admin controls and policy management; only step up to Enterprise if you specifically need SAML SSO or custom knowledge bases, and budget for the mandatory Enterprise Cloud add-on when you do.

Ease of use

This is Copilot’s clearest edge over editor-replacement tools: there’s genuinely nothing to learn about the environment itself, since it’s the environment you’re already in. The real learning curve is entirely about the AI layer — knowing when to accept an inline suggestion versus opening chat versus handing something to agent mode versus delegating a whole issue to the cloud agent, and increasingly, which model to pick for which kind of task now that the credit system makes that choice a cost decision as well as a quality one.

Who it’s actually for

Developers who want AI assistance without changing their editor or their team’s existing tooling — which, given Copilot’s reach across VS Code, JetBrains, Visual Studio, and more, covers most working developers by default. It’s also the more budget-friendly starting point if you’re evaluating whether AI coding tools are worth paying for at all. It’s a weaker fit if you specifically want the editor and agent built as one integrated product from scratch, or if your team’s threat model means “AI-reviewed” code needs to clear a higher bar than what the current agentic review feature has been independently shown to catch.