Cursor is an AI coding agent for real codebases: it plans and builds features, fixes bugs and reviews pull requests across desktop, CLI, Slack and GitHub. Cloud agents run in parallel on their own machines, with frontier models from OpenAI, Anthropic, Gemini and more.
Used this tool? Rate it
Used this tool? Rate it
Cursor is an AI coding agent built for working inside real codebases rather than answering questions about code in the abstract. The company's own framing has shifted decisively: where earlier descriptions called it the best way to code with AI, the site now leads with a single sentence — Cursor is your coding agent for building ambitious software. That change is not cosmetic. The product's centre of gravity has moved from an editor that suggests completions to a system where agents turn ideas into code and you spend your time reviewing decisions rather than typing them.
One practical note before the details: the legacy cursor.sh domain now redirects to cursor.com, so older links and directory listings still pointing at the former address reach the same product. In practice that means several surfaces working on the same repository. There is a desktop application, a command-line interface, a Slack integration, and GitHub pull request review. There are cloud agents that run on their own machines, work for extended periods, and hand back finished work. And there are automations that fire on a schedule or a trigger without anyone asking. The official documentation describes the day-to-day uses plainly: understand your codebase, plan and build features, fix bugs, review changes.
Some corporate context matters for anyone evaluating this seriously. Anysphere was incorporated in 2022 by four MIT students and grew unusually fast — annual recurring revenue crossed $100 million in January 2025 and surpassed $3 billion by early 2026. In 2026 the company was acquired: SpaceX acquired Anysphere in an all-stock transaction valued at $60 billion, making Cursor a wholly owned subsidiary. If you are choosing a tool your team will depend on for years, ownership and its implications for model strategy and roadmap belong in the evaluation, and the Limitations section returns to this.
There is a genuine free tier. Hobby is free, needs no credit card, and gives limited Agent requests with access to Composer — enough to judge the product, not enough for daily professional use.
On paid plans, Individual starts at $20 a month across Pro, Pro+ with three times the Pro agent limits, and Ultra with twenty times. Teams runs $40 per user per month and adds centralised billing, a team marketplace, Bugbot code review, team-wide privacy mode and SAML/OIDC SSO, with a Premium variant carrying five times the Standard agent limits. Enterprise is custom-quoted and adds pooled usage, invoice billing, SCIM seat management, repository and model access controls, audit logs and priority support.
The mechanic that matters most is not the sticker price but what sits behind it: every plan includes a set amount of model usage, and on-demand usage continues past that allowance billed in arrears. That means your actual monthly cost is a function of how hard you drive agents, not just which tier you chose. Heavy use of expensive frontier models on long-context tasks can exceed the subscription price substantially. Set a spend limit, and treat the subscription as a floor rather than a total. One further note on purchasing: subscriptions are sold only through cursor.com, with no authorised resellers, and the company states that accounts bought elsewhere may be suspended or terminated.
Usage-based costs are the most common complaint, and there is a documented history. In 2025 the $20 Pro plan moved from 500 fast responses plus unlimited slower ones to a usage model. Users reacted strongly: some users exhausted the $20 budget after just a few prompts, particularly on Claude models, and those who had not set spend limits faced unexpected charges. The company responded with an unusually direct apology from its CEO — "We recognize that we didn't handle this pricing rollout well and we're sorry" — rolled back limits and refunded affected users. The episode is worth knowing not because the company behaved badly afterwards but because it illustrates the structural point: with usage-based billing on expensive models, your costs are variable and you must manage them.
Ownership changed hands, and the implications are not yet settled. SpaceX acquired Anysphere in an all-stock transaction valued at $60 billion, making Cursor a wholly owned subsidiary. Analysts were not uniformly optimistic: Vital Knowledge's Adam Crisafulli observed that Grok has so far failed to make a dent in the frontier market led by Anthropic, OpenAI, Google and Meta, framing the acquisition partly as an attempt to strengthen the buyer's own AI position. For users, the open question is whether a tool now owned by a model developer will stay genuinely neutral about which models it recommends. Today the product still offers models from multiple competing providers; whether that persists is a reasonable thing to watch rather than assume.
AI systems in the loop can be confidently wrong, including the company's own. In April 2025, an AI help-desk agent named Sam invented a non-existent login policy, prompting cancellations before the company apologised and issued refunds. This is a useful reminder about the category as a whole: the same properties that make agents productive make them capable of asserting false things fluently. Review remains necessary, and that applies to support answers as much as to generated code.
Agent output still needs human review. Delegating a feature does not delegate responsibility. Large agent-produced diffs are genuinely harder to review carefully than small human ones, and the temptation to approve on the strength of a passing test suite is real. Teams adopting this workflow should strengthen review practice and test coverage, not relax them.
Independent structured ratings are scarce. Unlike consumer SaaS, this product has no substantial body of verified reviews on the major B2B rating platforms, and much of the material that ranks well in search is affiliate-driven pricing commentary rather than independent evaluation. That is a limitation of the available evidence, and it means claims about satisfaction — positive or negative — should be treated cautiously. The endorsements on the vendor's own homepage are real named people, but they are selected by the vendor.
Privacy protection is opt-in. With privacy mode enabled the company guarantees that code data is not used for training by it or its model providers — but the phrasing matters: it is a mode you enable. Teams handling sensitive code should turn it on at the admin level and verify it, rather than assuming it applies by default.
Regional and infrastructure constraints. The company states it does not use or maintain any infrastructure in China and does not use China-headquartered subprocessors. That is clarity worth having, but it is also a constraint for organisations that need in-country data handling there.
There is a free Hobby tier with no credit card required, offering limited Agent requests and access to Composer. Paid plans begin at $20 per month for individuals and $40 per user per month for teams.
Because each plan includes a fixed amount of model usage, and consumption beyond it continues as on-demand usage billed in arrears. Expensive models on long tasks consume the allowance quickly. Set a spend limit to make costs predictable.
In July 2025 the Pro plan changed from a request-based model to usage-based credits, and unclear communication left users with unexpected charges. The CEO publicly apologised, the company rolled back limits and refunded affected users. It is worth understanding because it reflects how usage-based billing behaves, not just a one-time mistake.
SpaceX, following an all-stock acquisition of parent company Anysphere valued at $60 billion, which closed in August 2026. Cursor operates as a wholly owned subsidiary.
Models from OpenAI, Anthropic, Gemini, SpaceXAI and Cursor's own Composer line, including Claude Opus 5, Claude Sonnet 5, Gemini 3.1 Pro, the GPT-5.6 family and Grok 4.5 and 4.6. Context windows range from 200k to 1M depending on the model.
Not when privacy mode is enabled — the company guarantees code data is not used for training by it or its model providers in that mode. It is a setting rather than the default, so enable it explicitly, and on teams have an admin apply it organisation-wide.
Yes. There is a CLI, a Slack integration, GitHub PR review, JetBrains and Xcode plugins, and TypeScript and Python SDKs. Cloud agents run independently of your machine entirely.
A SOC 2 Type II report is available on request through trust.cursor.com, alongside a commitment to at-least-annual third-party penetration testing. Enterprise plans add SSO, SCIM, access controls and audit logs.
No. Subscriptions are sold only through cursor.com, with no authorised resellers, and the company warns that accounts purchased elsewhere may be suspended or terminated.
Both. The Individual tiers are aimed at solo developers and the free Hobby tier allows evaluation without commitment; Teams and Enterprise add administration, shared context and governance that only matter at organisational scale.
