Ask a room of developers how they code with AI and three answers dominate: an agent in the terminal, an editor rebuilt around AI, or an assistant inside the editor they already had. Claude Code, Cursor and GitHub Copilot are the standard-bearers of each approach — less rivals in a single race than three different theories of where the AI should sit.
The numbers show all three theories winning at once. Claude Code passed a $2.5 billion revenue run-rate in February 2026. GitHub Copilot passed 50 million users in July 2026, with 4.7 million paid subscribers earlier that year. Cursor reportedly agreed to be acquired by SpaceX in June 2026, a deal that valued the editor layer as strategic territory.
Claude Code: the agent in the terminal
Claude Code works where DevOps has always lived: the command line. You describe a task, it explores the repository, plans, edits files, runs tests and reports back with a summary and a diff. The unit of work is the ticket, not the line.
Its strength is autonomy within a loop it can verify — tests pass or they don’t. Its demand on the user is managerial: clear task descriptions, conventions written down for the agent, and disciplined review of everything it produces. Developers who thrive with it describe managing a very fast, very literal junior team.
Cursor: the editor rebuilt around AI
Cursor starts from the opposite bank: keep the developer in the editor, but make the editor itself intelligent. Chat understands the whole codebase, edits span multiple files, and autocomplete predicts not just the next line but the next edit, including where your cursor is heading.
The feel is collaborative rather than delegated — you remain the author, with an extremely capable second pair of hands. For developers who think in code and want to stay in flow, that presence matters more than autonomy. The reported SpaceX acquisition suggests its future may tie into a much larger compute and distribution story, but the product’s identity is the editor itself.
Copilot: the assistant inside your editor
GitHub Copilot is the incumbent: an extension that brings completion, chat and increasingly agent features into VS Code, JetBrains and the rest of the tools teams already run. Its 50 million users reflect distribution as much as capability — it is the AI most developers met first, and the one their employers already pay for.
Copilot’s evolution mirrors the market’s: born as autocomplete, it has grown chat, multi-file editing and agent modes that narrow the gap with the other two. The differences today are of emphasis and default behaviour more than of kind.
Copilot’s quieter advantage is choice and governance. It lets users select between several frontier models for chat and agent tasks, and it arrives with the enterprise controls — policy management, audit logs, content exclusions — that large organisations already buy from GitHub. For a CIO, it is frequently the only one of the three that procurement can approve this quarter.
The pricing shapes differ in ways that matter at scale. Copilot is a per-seat add-on to an existing relationship, Cursor a per-seat editor subscription, and Claude Code usage-sensitive: seat tiers include an allowance, after which agent hours draw metered API billing. Two teams using “AI coding” can therefore have cost curves that look nothing alike.
And the three are converging. Copilot adds agent modes, Cursor adds background agents, Claude Code adds interfaces and review surfaces. Choosing today is really choosing a default centre of gravity — terminal, editor or extension — knowing the features will keep leaking across the borders.
Which one, and a hardware footnote
Match the tool to the work’s shape. Long, well-specified tasks with tests: Claude Code. Interactive building where you want your hands on the code: Cursor. A team standardising with minimal disruption: Copilot. Many professionals, tellingly, run two — an agent for the heavy lifts, an editor assistant for the flow.
The newest variable is location. Laptops built for local AI, like the RTX Spark machines in our coverage of AI laptops running agents on-device, ship with local agent tooling — a sign the three-way contest will soon include a fourth question: not just where the AI sits in your workflow, but whether it sits in the cloud at all.
0 Comments