I’ve built every version of this site in the same VS Code setup. Same theme, same six extensions, muscle memory over everything. So when a fork of VS Code kept showing up in developer forums as the tool that finally made AI coding stick, I paid attention. Could something that looks exactly like the editor I already live in be worth $20 a month on top of it?
That’s the question this review tackles. Cursor is built by Anysphere, it looks and feels like VS Code because it is one, and in late 2026 it’s probably the most-watched product in AI developer tooling. It has also changed its pricing three times in two years, and one of its flagship models turned into a small trust controversy this spring. Plenty to like here, and a couple of things that deserve hesitation. Let’s get into both.
What Cursor actually is
Under the hood, Cursor is VS Code. Extensions, keybindings, and your theme import in one click, which matters more than most reviews admit. The switching cost is nearly zero, and that’s a deliberate choice by the company. They knew developers rarely switch editors for features alone.
What’s new on top of the familiar shell is the AI layer. There’s Tab, a predictive completion model that suggests your next keystrokes. There’s the agent, which takes a plain-English request and makes multi-file edits while you watch it think. Cloud agents keep a task running even when you close the laptop. Bugbot reviews pull requests for bugs. MCP support lets you connect external tools and data sources. It’s a lot of surface area, and the interface does a decent job hiding most of it until you go looking.
Tab is the part nobody wants to give back
Read enough commentary on this product and a pattern appears: people arrive for the agent but stay for Tab. It predicts not just the end of your line but the next several, including edits to nearby code that haven’t happened yet. More than one long-term reviewer has described the same experience, their typing rhythm shifting because the suggestions land where their fingers were headed anyway.
Even the modest autocomplete I use daily inside VS Code has changed how I edit code, and Cursor’s version of this idea is the reason so many people tolerate everything else in the product. You can judge Tab yourself on the free Hobby tier, which asks for no credit card. That’s a smarter first step than any review, including this one.
Honesty first: how this review was built
Transparency note, because it belongs early: I haven’t run Cursor as my daily editor. Code on this site gets written in plain VS Code, and I’ll say plainly that the praise for Tab above is reconstructed from developer accounts, not my own muscle memory. What I did do: pulled every number and feature claim from Cursor’s own pricing pages and documentation as of October 2026, then compared notes across developer forums and the strongest review roundups published this year. The About page explains my full method for tools I haven’t lived in for weeks at a time. Where a claim traces back to a benchmark or a single roundup, I flag it. Reviews that blur that line don’t deserve your trust, and I’d rather keep mine.
Composer: Cursor’s own models, with a backstory
Early in 2026 the company shipped Composer 2, its second in-house coding model. Cursor’s published benchmarks put it ahead of Anthropic’s Claude Opus on Terminal-Bench 2.0, with input pricing around $0.50 per million tokens, roughly a tenth of what Opus costs. If that ratio holds in real workloads, it changes the math on agentic coding for anyone running agents all day. A later Composer update, per one September roundup I read while putting this together, reportedly pulled level with Opus 4.7 on coding benchmarks at similar savings.
Then came the wrinkle. Developers poking at Composer 2’s responses reported that it behaved like a distillation of Moonshot AI’s Kimi K2.5, a Chinese open-weights model. Whether that counts as normal industry practice, something shadier, or a gray zone in between got argued at length online, and Cursor’s public explanation didn’t fully settle the matter. A second wave of reviews in April concluded the model is legitimately fast and cheap, and that the attribution question is a governance issue more than a quality one.
I bring this up because of what it signals. You’re trusting this tool with your codebase and, increasingly, with your commit rights. A company that’s slippery about one piece of provenance earns watchfulness on the others. It doesn’t cancel out the engineering, but it belongs on the scales.
What you actually pay in 2026
Cursor’s pricing page, checked in October 2026, looks like this:
| Plan | Price | Who it’s for | The short version |
|---|---|---|---|
| Hobby | $0 | Tinkerers | Limited agent requests, Composer access, no card required |
| Pro | $20/mo | Most solo developers | Extended agent limits, frontier models, cloud agents, Bugbot on usage billing |
| Pro Plus | $60/mo | Heavy daily users | Bigger usage pools before overages kick in |
| Ultra | $200/mo | Agent-intensive workflows | The largest allowance; only worth it if agents run all day |
| Teams | $40/user/mo | Small to mid teams | Central billing, SSO, privacy mode, team rules and skills, usage analytics |
| Enterprise | Custom | Larger orgs | Pooled usage, audit logs, access controls, SCIM seat management |
Two notes on that table. One, Pro Plus and Ultra prices are the widely reported figures, consistent across roundups, while the pricing page presents them as tiers grouped under an Individual umbrella alongside Pro, which it lists explicitly at $20. Two, there’s also a Start plan aimed at developers in India that runs on Cursor’s own model pool at a lower price point. Annual billing brings the monthly rates down somewhat, which matters more at the Ultra level than anywhere else.
Billing is also where the loudest complaints live. Cursor moved to a credit-based system during 2025, overages bill on usage, and features like Bugbot and cloud agents can run up charges when agents chew through large tasks. Two usage pools split frontier models from Cursor’s own models, which hands you options but doubles the billing surface you need to understand. The sentence that repeats across forums goes something like: the product is great, the invoice is confusing. If you enable auto-run agents, set billing alerts before you unleash them.
How it stacks up against Copilot and the others
GitHub Copilot costs about half of Cursor Pro and its completions are solid. If what you want is smart autocomplete inside the editor you already have, Copilot is the right purchase. Where Cursor pulls ahead is the agent loop: repo-wide context, multi-file edits, terminal access, and taking a task from description to working code. Copilot has been catching up on all of that, and Microsoft’s distribution advantage is not small, but the agent experience on Cursor is more finished as of this writing.
Claude Code is a different shape entirely. It lives in your terminal and leans on Anthropic’s model line, which suits developers who think in shell commands. Windsurf competes closely on the fork-based editor idea and undercuts on price. None of these picks are wrong; they suit different working styles. If you want the longer argument, our guide to AI code assistants in 2026 works through that decision in detail.
Day one: what setup looks like
One thing worth knowing before you download: the first hour is intentionally boring. Cursor opens as VS Code, offers to import your current setup, and then behaves like the editor you had. Privacy mode is a toggle worth flipping during setup if code provenance matters to your clients, and I’d argue every review of this product should tell you where that switch lives before anything else. From there, the model picker in the chat panel is where most of the interesting decisions happen, because choosing between a frontier model and Composer is genuinely a cost-versus-speed call on the paid plans. Documentation for all of it is public and, unusually for a company moving this fast, readable.
Who should actually pay for this
- Solo developers and small shops shipping daily. A $20 editor pays for itself in one afternoon of saved yak-shaving.
- Agencies and consultancies. Juggling several unfamiliar codebases is where repo-aware chat earns its keep.
- Onboarding-heavy teams. Plain-English agents flatten the learning curve on a codebase nobody has documented.
- Teams of five to fifty. Shared rules, usage analytics, privacy mode, and SSO without building any of it yourself.
And who should hold off:
- Pure hobbyists, because the free Hobby tier may cover everything you do.
- Organizations already standardized on Copilot plus internal tooling; mid-flight editor migrations rarely go as smoothly as the demo suggests.
- Freelancers on tight budgets who mostly edit code rather than write it. Copilot’s completions at $10 may be plenty.
- Anyone in a regulated shop where auto-run agents and your compliance team still need to meet each other.
The downsides, stated plainly
- Billing complexity. Two usage pools, credits, and usage-based extras make invoices harder to predict than they should be.
- The provenance question. The Composer 2 attribution episode is small in practical terms and large in trust terms. Weigh it how you like, but know it happened.
- Electron overhead. It inherits VS Code’s RAM habits and adds models on top. Older laptops feel it, especially during long agent runs.
- Workflow lock-in. Rules files, custom skills, and agent configs accumulate over months. Leaving later costs more than starting does.
- Pricing history. Three pricing restructures in two years is a pattern. Whatever you sign up for, expect it to move.
Quick answers to the obvious questions
- Is there a free version? Yes. Hobby is $0 with limited agent requests and Composer access, no card required.
- What models can it run? In 2026, frontier models from OpenAI, Anthropic, Google, and others, plus Cursor’s own Composer line. Paid plans get two usage pools to balance intelligence against cost.
- Does my code get used for training? There’s a privacy mode that prevents that, and Teams plans can enforce it across the whole organization.
- Can I keep my VS Code setup? Yes, that’s the whole pitch. Extensions and settings import on first launch.
- Is it worth $20 over a free editor? If you write code most days, probably. If you edit config files twice a month, no.
Key takeaways
- Cursor Pro runs $20 a month, with a free Hobby tier below and $60, $200, and $40-per-seat tiers above it.
- Tab autocomplete is the feature users refuse to give back, and it’s testable for free.
- Composer 2 beats frontier models on Cursor’s own benchmarks at roughly a tenth of the token cost, with a live debate about its training provenance.
- The agent loop is the most complete of the mainstream editors right now; Copilot still wins on price for basic completions.
- Set billing alerts before enabling usage-based features. Surprise invoices are the most common complaint.
Affiliate disclosure: links to Cursor on this page are affiliate links. If you subscribe through one, this site may earn a commission at no extra cost to you. It doesn’t change the tradeoffs you just read.
If any of this has you curious, take the same path I’d hand a friend: start on Hobby, give the agent a week of real work on a real branch, and see how it feels. Then decide whether the Pro plan has earned its $20. Editors are a long-term relationship. This one has at least earned the first date.