Cursor

Rating reflects Cursor's documented plans and capabilities as of September 2026 plus our own sessions. Some vendor links on this site are affiliate links; they never affect a rating. See our disclosure and method.

What it does

Cursor is a code editor built on VS Code with AI woven through every part of it: Tab, the autocomplete that predicts your next edit; Agent, which takes a task, reads the codebase, edits files, runs commands and tests, and reports back; and, since Cursor 3 shipped on April 2, 2026, an Agents Window for running many of those agents in parallel across local folders, isolated git worktrees, cloud machines and remote SSH hosts. Agent Tabs show several sessions side by side, a /worktree command sandboxes an agent's changes off your branch, and /best-of-n runs the same task on several models at once so you can pick the outcome you like. Design Mode lets you click an element in a running web app and tell the agent what to change.

Cloud Agents, which Cursor renamed from Background Agents in its 2.0 release, run on Cursor's own virtual machines: they clone your repository, do the work, and hand back a pull request. You can start one from the desktop app, cursor.com, the iOS app, Slack, GitHub, Linear or the command line. Since August they can subscribe to events, so an agent watches the PR it opened, fixes CI and answers review comments without being asked, and a /goal command gives an agent a long-lived objective. On September 10 Cursor launched Projects, a coordinator agent that plans a large body of work, delegates to as many subagents as it needs, keeps a shared context folder that syncs to every machine, and can run on a schedule or watch a Slack channel. Self-hosted machines, added on September 2, keep tool execution inside your own network.

The models matter because Cursor charges by them. Every plan includes Composer, Cursor's own coding model: Composer 2 arrived in March 2026 and Composer 2.5 on May 18, both built on Moonshot's open Kimi K2.5 checkpoint and priced at $0.50 per million input tokens and $2.50 per million output. In April Cursor partnered with SpaceXAI to train Grok 4.5, released July 8, and SpaceX completed its acquisition of Cursor on August 14; Grok 4.6 followed that week. The pricing page now advertises "generous limits for Grok" on paid plans. Alongside those sit the frontier models from Anthropic, OpenAI and Google, billed against a monthly allowance at their API rates.

Where it wins

• Agent management. No other editor makes running five agents at once feel normal. Worktrees keep them from colliding, the Agents Window shows what each is doing, and cloud agents let you close the laptop and get a PR in the morning.
• Model choice. Composer for fast, cheap iteration; Grok, Claude, GPT and Gemini models when the task needs it; /best-of-n when you are not sure. Claude Code gives you one vendor's models; Cursor gives you all of them.
• Reviewability. Every agent change lands as a diff in an editor you already know, with Bugbot's automated review available on usage billing. For teams that need to see what the agent did before it merges, that beats a terminal transcript.
• Composer's price. At $0.50 and $2.50 per million tokens, Composer 2.5 is a fraction of the cost of frontier models, and Cursor's benchmarks put it close to last year's Opus and GPT models for everyday agent work. It is the reason the $20 plan stretches as far as it does.
• Enterprise controls. Team-wide Privacy Mode, SSO, SCIM, repository and model access controls, audit logs, self-hosted execution and an AI code tracking API. Cursor holds SOC 2 Type II, ISO 27001 and ISO 42001.

Where it loses

• The meter. The pricing page publishes tiers and multipliers, not limits: Pro is "extended limits", Pro+ is 3x Pro, Ultra is 20x Pro. Every plan includes a set amount of model usage and then bills on demand at model rates. Agent-heavy developers routinely find that $20 turns into $60 or $200, and Cursor itself now recommends Pro+ for daily agent users.
• Frontier work still goes elsewhere. Composer is fast, but for the hardest refactors and debugging, Anthropic's Opus 5 and Fable 5.1 and OpenAI's top models still win, and those are the expensive ones on Cursor's meter. Claude Code on a $20 Claude Pro plan gives you Opus 5 with a flat fee.
• Privacy Mode is a setting, not a default. With it off, your code, prompts and editor actions may be stored and used for training. Individuals must check it; only Teams and Enterprise can enforce it.
• Ownership. Cursor is now a SpaceX company that trains models with SpaceXAI and pushes Grok on its pricing page. If model neutrality was part of why you chose it, watch how the defaults move.
• Complexity. Agents Window, worktrees, cloud agents, Projects, subscriptions, custom modes, self-hosted pools: the product has grown faster than most people's mental model of it. Beginners should start with Tab and a single Agent.

Plans and pricing, September 2026

Enterprise is custom-priced and adds pooled usage, invoicing, SCIM, repository and model access controls, audit logs and priority support. Independent trackers report about 20% off on annual billing; Cursor's own pages do not print annual rates, so confirm at checkout. An India-only Start plan at 649 rupees a month covers Cursor's models only. Every paid plan includes a set amount of model usage, with on-demand usage billed in arrears at model rates once it runs out.

Score breakdown

Headline rating weights Capability 30%, Ease of use 20%, Value 20%, Trust 15% and Ecosystem 15%, the same weights we use for every card.

Alternatives

• Claude and Claude Code if you want Anthropic's strongest models on a flat fee and prefer the terminal to an editor.
• ChatGPT if you want Codex included on every plan and a general assistant in the same subscription.
• Gemini if you are already paying for Google AI Pro, which bundles the Jules coding agent and Antigravity access.

Quick answers

Is Cursor free?

There is a free Hobby plan with no credit card required. It gives you the editor, limited Agent requests and access to Composer, Cursor's own model. Pro is $20 a month, Pro+ is $60 with three times the Pro agent limits, and Ultra is $200 with twenty times.

Does Cursor train on my code?

Not if Privacy Mode is on. With it enabled Cursor guarantees your code is not used for training by Cursor or its model providers, and it holds zero-retention agreements with them. With Privacy Mode off, Cursor may store and use your code, prompts and editor actions to train its models. Check the setting; Teams can enforce it for everyone.

Cursor or Claude Code?

Cursor if you want an editor with agents in it, several models to choose from, and a visual place to review what they did. Claude Code if you live in the terminal and want one very strong model that runs autonomously for hours. Many developers pay for both: Cursor Pro at $20 and Claude Pro at $20.

More cards

• 
  Tool cards

  Claude

  The writer's model, with Claude Code in the $20 plan.

• 
  Tool cards

  ChatGPT

  The model-first assistant, with Codex on every plan.

• 
  Guides

  Build your first AI stack

  Which tools to pay for, in what order, for what you do.