Last updated: 2026-07-01
Cursor is Anysphere's AI code editor surpassing $2B ARR at $29.3B valuation (Feb 2026), featuring Composer 2.5 agent for multi-file edits and 88% SWE-bench via integrated models.
About Cursor
Cursor is an AI-powered code editor built by Anysphere (founded 2022, Cambridge MA), a fork of VS Code that replaces the extension layer with deep, system-level AI integration. As of early 2026, it is used by 64% of Fortune 500 companies — including 20,000+ developers at Salesforce alone — and reached $2 billion in annualized revenue by February 2026 on a trajectory to exceed $6 billion by year-end. Every VS Code theme, keybinding, and extension works in Cursor without reconfiguration. The core product advantage over Copilot and other AI extensions is architectural access: by forking VS Code rather than running as an extension, Cursor controls the editor's rendering pipeline, file system, and extension host. This enables the Shadow Workspace for semantic codebase indexing, speculative suggestions that render in real time, and Background Agents running in isolated VMs on Cursor's cloud infrastructure. Agent Mode lets developers describe a task in plain language; the AI plans, edits multiple files, runs terminal commands, installs packages, and iterates on failures without manual prompting. Benchmarks show Tab completions averaging 200ms latency, roughly 50ms faster than GitHub Copilot. In late 2025 and early 2026, Cursor expanded beyond the IDE. Cloud Agents run autonomously on Cursor's infrastructure and can be triggered from a browser, mobile device, or Slack — editing code, running tests, and filing pull requests while the developer is away from the desktop. Cursor v3.0 introduced Background Agents supporting up to 10 parallel workers per user and Composer 2.0 for coordinated multi-file changes trained with 20x scaled reinforcement learning. Native MCP (Model Context Protocol) integration connects agents directly to GitHub, Linear, Jira, Postgres, Supabase, and Slack without context-switching. Pricing spans five tiers as of June 2025: Hobby (free, limited completions and agent requests), Pro at $20/month with a $20 monthly credit pool, Pro+ at $60/month, Ultra at $200/month, and Teams at $40/user/month with invoice billing, SCIM seat management, an AI code tracking API, and granular admin controls. Annual billing saves 20% across all paid tiers. Privacy Mode with Zero Data Retention terms is available on all plans; neither Cursor nor its LLM providers train on customer code in Privacy Mode. Cursor's primary weakness is context at scale: large codebases exceeding tens of thousands of lines force the AI to make tradeoffs on what to include in context, sometimes choosing wrong. AI operations increase CPU and memory usage by up to 30%, and latency spikes occur in roughly 25% of large codebase tasks. The tool only supports GitHub for version control integration natively; GitLab users must use workarounds. Anysphere does not currently sign HIPAA BAAs, making Cursor unsuitable for healthcare organizations handling PHI.
Pricing
Hobby (free) with limited completions and agent requests. Pro at $20/month ($16/month annual) with a $20 monthly credit pool and unlimited Tab completions. Pro+ at $60/month. Ultra at $200/month. Teams Standard at $40/user/month (invoice billing, SCIM, admin controls). Teams Premium at $120/user/month ($96/user annual) for 5x Standard usage — designed for heavy agent users. Annual billing saves 20% across all paid tiers. Auto mode is unlimited on all paid plans; credits apply only to manually selected frontier models. Composer 2.5 credit-based add-on: 1 credit per agentic action (multi-file edit with 3+ files), $0.02-$0.08 per action depending on model.
Key Features
- Agent Mode: Describes a task in natural language and Cursor autonomously edits multiple files, runs terminal commands, installs packages, and iterates on test failures until the task is complete.
- Tab Completion: Project-aware multi-line autocomplete that predicts the next 3 lines based on surrounding context, running at roughly 200ms latency, approximately 50ms faster than GitHub Copilot.
- Cursor Automations (v3.8, June 2026): Always-on background agents that automate repetitive tasks on schedule. Use /automate to describe a workflow in plain language; agents respond to GitHub and Slack triggers, and cloud agents can now use computer use to produce demos or artifacts of their work.
- Cloud Agents & Fast Environments: Agents run on Cursor's cloud infrastructure rather than the local machine. Version 3.7 (June 2026) delivers cloud environments that spin up in under 10 minutes via reusable snapshots, with /in-cloud for isolated subagents and seamless local-to-cloud session handoff.
- Composer 2.0 (Multi-File Editing): Edits 5 to 20 files simultaneously with coordinated changes, presenting visual diffs to review file by file, trained with 20x scaled reinforcement learning over the v1 model.
- Native MCP Integration: Connects agents directly to GitHub, Linear, Jira, Postgres, Supabase, and Slack via Model Context Protocol, letting agents open issues, query schemas, and file PRs without leaving the IDE.
- Privacy Mode with Zero Data Retention: Enabling Privacy Mode activates contractual ZDR terms with all LLM providers so no customer code is stored or used for training, available on all plans including Hobby.
- Cursor Router (Auto Mode): Cursor Router launched July 22, 2026, an intelligent model router that classifies each request by complexity and domain and routes it to the right model automatically. Trained on 600,000+ live requests, Cursor says it delivers frontier-quality results at roughly 60% lower cost, with Intelligence, Balance, and Cost modes available across desktop, web, iOS, CLI, and SDK.
Pros
- 64% of Fortune 500 companies use Cursor as of early 2026, with 20,000+ developers at Salesforce alone, making it the most widely adopted AI code editor in enterprise.
- Tab completions average 200ms, roughly 50ms faster than GitHub Copilot, and Agent Mode reduces boilerplate and feature scaffolding tasks from an afternoon to minutes.
- Every VS Code extension, theme, and keybinding migrates in one click, eliminating the learning curve and tool switching cost entirely.
Cons
- Large codebases exceeding tens of thousands of lines hit context limits, forcing the AI to choose what to include and sometimes choosing wrong, producing confident but incorrect edits.
- Native version control integration supports only GitHub; GitLab users must rely on unofficial workarounds, which is a significant gap for the majority of enterprise teams on GitLab.
- Anysphere does not sign HIPAA Business Associate Agreements, making Cursor unsuitable for any healthcare organization handling Protected Health Information.
Frequently Asked Questions
What is Cursor and who built it?
Cursor is an AI code editor built on a fork of Visual Studio Code, developed by Anysphere, Inc., a company founded in 2022 by MIT students Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger. By February 2026, Cursor's annualized revenue surpassed $2 billion, with some estimates reaching closer to $3 billion by April 2026, making it one of the fastest software products ever to reach $1 billion in revenue, in roughly 18 months. Anysphere closed a $900 million Series C at a $9.9 billion valuation in mid-2025, then a Series D at a $29.3 billion valuation in November 2025, and reports in April 2026 indicated talks for a new round near $50-60 billion led by a16z and Thrive Capital. About two in three Fortune 500 companies now use Cursor, and the company reported over 1 million daily active users in 2025.
How much does Cursor cost?
Cursor has six pricing tiers. Hobby is free with limited Tab completions, limited agent requests, and Privacy Mode. Pro costs $20/month ($16/month billed annually) and adds a $20 monthly credit pool and unlimited Tab completions. Pro+ costs $60/month for a larger credit pool, and Ultra costs $200/month for the highest usage limits. Teams Standard costs $40/user/month with invoice billing, SCIM, and admin controls. Teams Premium costs $120/user/month ($96/user annual) for 5x the Standard usage, with separate usage pools for Cursor-native models (Composer, Auto) and third-party API models (Claude, GPT, Gemini) — designed for heavy agent users who would otherwise hit limits unpredictably. Annual billing saves 20% on all paid tiers. On the model side, Cursor's own Composer 2.5 (released May 18, 2026) costs $0.50 per million input tokens and $2.50 per million output tokens in its standard form, or $3.00/$15.00 per million tokens for the faster variant.
What does Cursor do that other code editors don't?
Cursor's Agent Mode autonomously edits multiple files, runs terminal commands, installs packages, and iterates on test failures from a single natural-language instruction. Composer 2.0 and 2.5 coordinate edits across 5 to 20 files at once with visual file-by-file diffs; Composer 2.5 (May 18, 2026) raised SWE-Bench Multilingual from 73.7% to 79.8% and Terminal-Bench from 61.7% to 69.3% over Composer 2. Cloud and Background Agents run on Cursor's own infrastructure rather than the local machine, supporting up to 10 parallel agents triggered from a browser, mobile device, or Slack. Tab completion runs at roughly 200ms latency, about 50ms faster than GitHub Copilot, and native MCP integration connects agents directly to GitHub, Linear, Jira, Postgres, Supabase, and Slack. Privacy Mode, which provides contractual Zero Data Retention with every underlying LLM provider, is included on every plan, including the free Hobby tier.
How does Cursor compare to GitHub Copilot and Claude Code?
Cursor's Tab completion runs roughly 50ms faster than GitHub Copilot at about 200ms latency, and its Agent Mode and Composer 2.5 handle multi-file, repo-wide edits with autonomous test iteration that goes well beyond Copilot's largely completion-focused model. Cursor is model-agnostic, routing tasks to Claude Sonnet, Claude Opus, GPT-4o/4.1, or Gemini 2.5 Pro, while also offering its own Composer 2.5 model at $0.50/M input and $2.50/M output tokens. Claude Code, Anthropic's terminal-based coding agent, reached around $2.5 billion in annualized revenue by February 2026 and is often favored for deep reasoning on large refactors, while Cursor's full IDE replacement and $2 billion+ ARR reflect its strength as an everyday editor for individuals and large engineering teams alike. Both offer free entry tiers and a $20/month plan, with Cursor adding Pro+ ($60/month) and Ultra ($200/month) for heavier agentic workloads.
Does Cursor have a free tier?
Yes. Cursor's Hobby tier is free and includes limited Tab completions, limited agent requests, and Privacy Mode (contractual Zero Data Retention with underlying LLM providers), even without a paid plan. It does not include the $20 monthly credit pool, unlimited Tab completions, or the higher usage limits of Pro+ and Ultra. Auto mode, where Cursor automatically selects an appropriate model, is unlimited on paid plans but is more limited on Hobby. For unlimited Tab completions and a working credit pool for agentic tasks, most individual developers move to Pro at $20/month ($16/month billed annually).
Who is Cursor best for, and who should be cautious?
Cursor suits individual developers and engineering teams doing repo-wide, multi-file work, since Agent Mode and Composer 2.5 can edit 5-20 files at once, run tests, and iterate on failures; about two in three Fortune 500 companies already use it. Teams needing SCIM, invoice billing, and admin controls should use the $40/user/month Teams plan. Developers working in very large repositories (500,000+ lines of code) should be aware indexing can occasionally miss files or make incorrect cross-file connections, and should rely on .cursorignore and explicit @-file references. Organizations on networks with proxies that block HTTP/2, such as Zscaler, may see degraded AI streaming and should request an exception, since Cursor depends on HTTP/2 for chat, completions, and agent responses.
What is the Cursor code reversion bug and how do I avoid it?
In March 2026, Cursor's team confirmed that code changes could silently revert due to file-locking conflicts between three sources: the Agent Review Tab, cloud sync tools like OneDrive or iCloud, and the editor's Format On Save feature. To avoid it, close the Agent Review Tab before using 'Fix in Chat', exclude your project folder from any cloud sync service, and disable Format On Save while running Agent Mode. For general problems such as startup crashes, a frozen chat panel, or broken autocomplete, clearing Cursor's local cache resolves most cases. Before troubleshooting locally, check status.cursor.com to rule out a platform-wide outage, since AI streaming issues caused by HTTP/2-blocking proxies (like Zscaler) can look similar to a local cache problem.
Does Cursor train on my code, and is it compliant for regulated industries?
Cursor does not train its models on customer code by default. Enabling Privacy Mode, available on every plan including the free Hobby tier, activates contractual Zero Data Retention terms with all underlying LLM providers (Anthropic, OpenAI, Google), so none of those providers store or train on the code either. Cursor maintains SOC 2 Type II, GDPR, and CCPA compliance, but Anysphere does not advertise HIPAA compliance, so healthcare organizations handling protected health information should isolate Cursor from PHI-containing repositories or evaluate alternatives. Cursor's underlying LLM providers have been subject to EU AI Act general-purpose AI model obligations since August 2025, which indirectly extends to models accessed through Cursor.
Top Alternatives
- Windsurf: Pick Cursor for Composer 2.5 maturity and $2B ARR validation; pick Windsurf for Cascade agent and free tier.
- GitHub Copilot: Pick Cursor for AI-native IDE and Composer agent; pick Copilot for VS Code integration and enterprise trust.
- VS Code: Pick Cursor for built-in AI agent; pick VS Code + Copilot for familiar editor + extension.
HokAI guides covering Cursor
- Cursor Composer 2.5 vs. Claude Code: Which AI Coding Agent Should You Use in 2026?: Cursor Composer 2.5 vs Claude Code compared on 2026 pricing, benchmarks, and context windows now that Claude Code defaults to Sonnet 5 and Opus 5, not Opus 4.6.
- xAI Spent $60 Billion on Cursor. Grok Build Now Runs Two Coding Models.: SpaceX's $60B Cursor deal means xAI's Grok Build now runs Composer 2.5 and Grok 4.5 side by side, with no vendor benchmark comparing the two coding engines.
- Claude Max Used to Die by Wednesday. Anthropic's Fix Expires August 19.: Anthropic doubled Claude Code's limits in May and boosted them 50% more since. That extra 50% is a promotion, extended four times, due to end August 19.
- Netlify Stopped Charging by Seat Because Agents Don't Have One: Netlify killed per-seat pricing on April 14, 2026, folding unlimited AI agents into one flat $20 Pro plan, while Vercel and Replit still bill per seat.
- How to Build an AI-Powered App Without Being a Developer: Lovable hit $500M ARR in ten months. Real prices across six no-code and prompt-to-app builders, plus the security check most founders skip before launch.
- How to Automate Repetitive Work With AI (Without Writing Code): Zapier now lists 9,000-plus apps, Make bills in credits, and Notion added external AI agents. Updated 2026 numbers for choosing a no-code automation platform.