Anysphere · Cursor
Cursor: a meter in dollars, and a tank with no size
An AI code editor with an agent, a terminal client and cloud agents on top, sold since October 2026 as one plan with Grok and X, from a weekly pool sized only in multiples.
Cursor is a proprietary AI code editor from Anysphere, now part of SpaceX, with a CLI, cloud agents and a pull-request reviewer attached. Since October 2026 new customers buy it as The X Subscription: a Free tier, four paid tiers, team seats and an enterprise contract. Its own models draw on one weekly pool shared with Grok; third-party models bill at list price from the first request.
What Cursor actually is
A fork of VS Code with an AI layer built into it: an agent that plans and edits across files and runs commands, a Tab autocomplete that is the part many people switched for, and a chat panel. Around the editor sit a command-line client, cloud agents you can start from the web or an iPhone, a pull-request reviewer called Bugbot, SDKs for TypeScript and Python, and connectors into Slack, Linear, Jira, GitHub and JetBrains IDEs over the Agent Client Protocol. The queue filed it as a coding agent, and — for the second wave running — that label needs no correction.
The company behind it is Anysphere, Inc., and since 14 August 2026 Anysphere belongs to SpaceX, which completed an acquisition that began as a model-training partnership in April. That is why three of the four models in Cursor's own pool are Grok models, described as jointly trained with SpaceXSI. The client is closed: the public GitHub repository is an issue tracker with no licence, and the terms forbid derivative works. So this is not a free program with a paid service beside it, as with most tools in this section — it is a product, sold in tiers, with a free tier at the bottom.
What makes it worth a close read is the metering, which changed on 12 October 2026. New customers now buy The X Subscription, one plan for Cursor, Grok and X, sold by X Subscription Corp, and Cursor's own models draw on a weekly pool shared with the Grok apps and Grok Bot. Third-party models are paid from prepaid credit at each model's list price from the first request, and the vendor's pricing policy ties those rates to each model maker's list. That turns Cursor's rate card into a claim anyone can check, and we checked it: of the forty-one third-party rows our snapshot also carries, forty match to the cent and one does not. What the vendor does not publish is the other half — how much a pool holds. Tiers are stated as multiples of Plus, and Plus has no stated size.
We have not run it. Even the free tier needs an account, and this desk does not create accounts or enter passwords. Everything on these three pages is the vendor's documentation, terms, pricing policy and rate card, the X Subscription documentation and bundled terms, read on the date at the top, plus archived copies of the same pages and the arithmetic our own price snapshot allows. Where our own published comparisons had drifted from what the vendor says today, we corrected them the same day.
Seven routes: four paid tiers, team seats, an enterprise contract and your own key, with a Free tier beneath them. The prices are clear. The pool each one buys is stated only as a multiple of another pool nobody has sized.
| Route | Bought as | Price |
|---|---|---|
| Plus | Subscription | $8/mo $7/mo billed annually |
| Premium | Subscription | $30/mo $25/mo billed annually |
| Super | Subscription | $100/mo $84/mo billed annually |
| Ultra | Subscription | $200/mo $167/mo billed annually |
| Teams | Per seat | $30/seat/mo |
| Enterprise | Per seat | Custom |
| Your own API key | API key | At model rates |
When the honest answer is use something else
Seven published pairs include Cursor, more than any other tool we cover. These are the six with the most searches; the seventh, against Cline, is linked from the alternatives page. Each one lives in /compare/ and is written there rather than twice.
| Pair | What it decides | Verified |
|---|---|---|
| Cursor vs Claude Code | Editor against terminal. On the neutral index Claude Code scores higher and costs several times more per task; the pair decides which of those two numbers your work is sensitive to. | 2026-08-13 |
| Windsurf vs Cursor | Two VS Code forks on the same frontier models. Windsurf, now Devin Desktop, wins on autocomplete terms and deployment reach; Cursor on the agent stack and a middle tier. | 2026-07-27 |
| Cursor vs VS Code | The same editor underneath. What the subscription buys is the AI layer, and the pair asks whether it beats a free editor plus your own key. | 2026-08-13 |
| Cursor vs GitHub Copilot | A published allowance and every IDE, against the deeper editor loop. The pair lays the two allowance systems side by side, row by row. | 2026-07-27 |
| Codex vs Cursor | The stronger agent against the cheaper one, by wide margins in both directions — and, with the model held constant, how much of the gap is the harness. | 2026-08-13 |
| Antigravity vs Cursor | Free and metered against paid and priced. Google's harness scores higher on the index; the pair is about what the free tier stops you doing. | 2026-08-31 |
Four reasons people look past Cursor, and a different shortlist behind each one: the unsized allowance, the data path, wanting a terminal or your own IDE, and the price of a team seat. Organised by why you are going rather than by rank.
Frequently asked.
What people search next.
Q · 01 Is Cursor free? +
There is a Free tier on The X Subscription: the agent and chat with limited usage, a limit not stated as a number, and no third-party models. It needs an account, and the old free plan, Hobby, maps to it. There is no paid trial documented for the editor itself; the only trial in the documentation is a usage credit for Grok Bot.
Q · 02 How much usage does a Cursor plan include? +
The vendor does not say. Cursor's own models draw on one weekly pool per seat, shared with Grok and Grok Bot, and the tiers are sold as 4x, 15x and 50x of a Plus pool with no stated size. Third-party models have no pool: they are paid from prepaid credit at list price from the first request. Legacy plans keep two monthly pools, and until late August the documentation printed the third-party allowance of each; that column was removed between 24 August and 3 September 2026.
Q · 03 Does Cursor train on my code? +
Only if you have agreed, according to the terms — and the data-use overview says that turning Privacy Mode off is how that happens. Privacy Mode is documented as on by default for teams and enterprise. For personal accounts the default is not stated anywhere we could find, so the setting is worth checking before the first prompt. Two exceptions hold even with it on: requests with your own API key follow your provider's policy, and a few models — Anthropic's Fable line among them — keep inputs at their maker for safety review.
Q · 04 Who owns Cursor? +
SpaceX. Anysphere, Inc., the company that makes Cursor, announced on 14 August 2026 that SpaceX had completed its acquisition, a process that began with a model-training partnership in April. The Grok 4.5, 4.6 and 4.7 models in Cursor's own pool are described as jointly trained by Cursor and SpaceXSI. Since October 2026 Cursor, Grok and Grok Bot are sold together on The X Subscription by X Subscription Corp, while the site footer still says Anysphere.
Q · 05 Is there a cheaper Cursor plan in India? +
There was. Cursor Start, priced in rupees and sold only in India, is now a legacy plan: it keeps working for people who hold it, and the vendor's own mapping points it at Plus. We list Plus at its dollar price because that is the one the vendor publishes.
Q · 06 Can I use Cursor with my own API key? +
Yes from Premium up, for chat models from OpenAI, Anthropic, Google, Azure OpenAI and AWS Bedrock; Plus excludes it, and Tab completion keeps using Cursor's models. On individual tiers your provider bills the tokens and nothing is drawn from your pool. On Teams and Enterprise eligible requests, including those on your own key, carry a per-token platform fee. Either way the request passes through Cursor's servers, and the vendor's zero-retention agreements do not cover it.