Cursor hit 1 million paying developers sometime in March 2026 — Anysphere hasn’t announced the exact figure, but the engineering job market is talking about it like it’s already settled. What’s interesting isn’t the number. It’s how fast Cursor displaced Copilot as the default answer when a senior engineer asks a junior “what should I install first?”
I moved off VS Code with Copilot as my daily driver in February. Not reluctantly — enthusiastically. This is the first dev tool since Git that changed how I work in a way I notice every single day.
What Cursor actually is
Cursor is a fork of VS Code with AI deeply woven into the editor, not bolted on. That’s the entire pitch, and it matters more than it sounds.
The distinction between “woven in” and “bolted on” is the difference between asking Copilot for a line completion and telling Cursor “refactor the authentication flow to use JWT instead of session cookies across every file that touches it, update the tests, and flag anything you’re uncertain about.” The first is autocomplete. The second is delegation.
If you want the broader landscape, the best AI tools in 2026 guide covers where Cursor fits against Copilot, Replit, and the rest. What I want to do here is go deeper than any feature list would.
The agent mode is where the gap opens
Cursor’s Agent mode lets you give the IDE a goal rather than a prompt. You type what you want, it reads the relevant files, plans an approach, makes the edits across multiple files, runs the test suite, and reports back.
I used this last month to migrate a 14,000-line TypeScript codebase from Yup to Zod for schema validation. That’s the kind of task where the mechanical work is boring but every edge case matters. I wrote a single prompt describing the goal, reviewed the plan Cursor produced, accepted it with two minor changes, and watched it execute. Total wall-clock time: 47 minutes, of which I was actively engaged for maybe 15.
Manual equivalent: a team member spent 1 week on the same migration in a different codebase at my previous company. Different complexity profile, but still — the delta is not subtle.
What Agent mode does not do well is tasks that require genuine architectural judgment. I asked it once to “simplify this authentication flow” without specifying how. It rewrote the flow to be shorter, but traded off a security property I hadn’t mentioned but clearly wanted. When I catch it doing that kind of thing I add it to my own mental list of prompts that require more specification.
How the models work under the hood
Cursor doesn’t train its own frontier models. It routes your queries to Anthropic’s Claude (Sonnet 4.6 and Opus 4.7) and OpenAI’s GPT-5, depending on the task and your preference.
This is worth understanding before you pay. Cursor’s Pro plan bundles “unlimited” usage with fair-use caps on the premium models. You get generous access to Sonnet 4.6 — which is genuinely where the magic happens for most coding tasks — and metered access to Opus 4.7 for harder reasoning.
Opus 4.7’s release in April 2026 made a real difference to Cursor’s output quality for complex refactoring. If you want the full picture on which Claude model does what, the Claude 2026 guide covers it.
The fact that you can pick your underlying model matters when you have preferences. I run Sonnet 4.6 by default because it’s fast and correct for 90% of what I ask. For anything involving concurrency, distributed systems, or cryptography, I manually switch to Opus 4.7 and accept the slight latency.
Pricing and the Business tier question
Hobby: Free. 2,000 completions/month. Good enough for weekend projects, not for daily work.
Pro: $20/month. Effectively unlimited tab completions, generous premium model access, Agent mode, Composer (multi-file editing). This is what I pay.
Business: $40/user/month. Team privacy controls, centralized billing, SSO, zero data retention. For teams over ~10 developers, this starts making sense.
The honest question for teams: is Cursor Business worth 2x the price of Copilot Business ($19/user)? My answer is “usually yes” because the Agent mode productivity compounds for senior engineers. For a team of ten, you’re paying $2,400/year extra per developer. One Agent-assisted migration or one complex refactor that would have taken a week pays for that several times over. If your team writes mostly CRUD apps with small scope, Copilot is fine.
Where Cursor is still rough
Three areas where I genuinely wish it worked better.
Indexing large monorepos is slower than it should be. A 200,000-file monorepo takes Cursor about 8 minutes to reindex after a branch switch. VS Code with Copilot is effectively instant because it’s not trying to understand the full context. This is a tradeoff, not a bug, but it shows.
Agent mode occasionally loses track of its own changes. About once every twenty Agent-mode runs, I’ve seen it “forget” it had already edited a file and make a second, conflicting edit. The fix is usually to reject the session and start over with a more specific prompt. Annoying, not blocking.
The “Composer” feature still feels underbaked compared to Agent mode. It’s supposed to be the lighter-weight, interactive version for multi-file work, but I keep defaulting back to full Agent mode. Either Cursor needs to merge them or commit harder to differentiating them.
Where Cursor wins vs where Copilot still wins
I wrote a longer breakdown in the Cursor vs GitHub Copilot comparison, but here’s the short version.
Cursor wins when you’re an individual developer or small team writing code that requires deep context. Refactoring, migrations, net-new features with real architectural decisions, agent-assisted work. It wins on model quality because it can use Claude Opus 4.7 and GPT-5 side by side.
Copilot wins when you need frictionless enterprise rollout. If your org already runs GitHub Enterprise, Copilot’s admin tooling, audit logs, and compliance story beat Cursor’s. It also wins on IDE breadth — Copilot works in JetBrains, Vim, Neovim, and a dozen other environments where Cursor doesn’t have a presence.
If you’re on Replit for other reasons, Replit’s AI story is a different value proposition entirely — browser-based, deploy-in-one-click, less control but zero setup.
When to use Cursor, when not to
Use Cursor if: you’re a professional developer, you work in TypeScript/JavaScript/Python/Go (where Cursor shines), you’re comfortable reviewing AI-generated code before accepting, and you want a real shot at offloading mechanical refactoring work.
Skip Cursor if: your IDE of choice is a JetBrains product you can’t leave, you’re on a team that’s standardized on Copilot for compliance reasons, or you mostly write code where Agent mode’s power is overkill (simple CRUD, quick scripts).
The honest verdict
Cursor isn’t perfect, and the competition isn’t standing still. Zed has announced its own agent mode for later this year. Microsoft is clearly going to eventually respond with Copilot Workspace as a serious daily driver. The gap Cursor has right now — the combination of agent capability plus model choice plus actual code understanding — is not guaranteed to last.
But in April 2026, if you’re paying for exactly one coding tool, Cursor Pro at $20/month is the one. Nothing else currently delivers the same productivity compounding for individual developers.
I’m not going back to VS Code with Copilot. I miss fewer things than I expected. The things I do miss — faster indexing, a few JetBrains-specific refactors — aren’t worth trading the Agent mode back for.