Cursor

Overview

Cursor (developed by Anysphere) is one of the largest products in the AI coding track by valuation and scale in 2026: reaching **1 million users + 360,000 paid customers** in 16 months, with **annualized revenue of approximately $2 billion**. **In May, SpaceX locked in a $6 billion acquisition option** (preempting the a16z consortium that was raising $2 billion at a $50 billion valuation), with valuation multiplying 24 times in 16 months.

The product itself is an **"AI-native"** code editor, deeply modified based on VS Code—Tab completion, Cmd+K inline editing, Composer multi-file refactoring, Agent mode, and terminal execution are all native capabilities, not plugins.

**Latest version Composer 2.5** (launched on 2026/05/18): Self-developed model upgrade, base model based on **Moonshot's open-source Kimi K2.5** (trillion-parameter MoE, approximately 32 billion activated), with 85% of compute used for post-training + text feedback reinforcement learning + 25x synthetic data. **SWE-Bench Multilingual 79.8%** (at launch, only 0.7 percentage points behind Opus 4.7); **CursorBench v3.1 surpasses Opus 4.7** (63.2% vs 61.6%); price only **1/10** of the contemporary Opus flagship ($0.50/$2.50 per M tokens). This release proactively disclosed the Kimi K2.5 base model source, significantly improving transparency.

**2026/7/9 SpaceX×Cursor relationship materialized into a specific product**: Both parties jointly trained and released **Grok 4.5** (1.5T parameter V9 base model, incorporating trillions of real Cursor developer interaction data), and it has been natively integrated into the Cursor editor, becoming a new cost-effective model option—Token consumption for completing SWE-bench Pro tasks is only 1/4 of Opus 4.8. **Note**: Anthropic released Opus 4.8 on 5/28 and the more cost-effective Sonnet 5 on 6/30. The comparison benchmarks at the time of Composer 2.5's launch are no longer the latest flagships, but the price advantage logic still holds.

Key Features

• Composer 2.5 Self-developed Coding Model: Upgraded on 5/18, base model Kimi K2.5 trillion-parameter MoE. SWE-Bench Multilingual 79.8%, CursorBench v3.1 surpasses Opus 4.7, price only 1/10 of Opus 4.7
• Tab Smart Completion: Multi-line prediction, accept step by step with Tab, supports project-level context
• Cmd+K Inline Editing: Select code, describe modification intent in natural language, AI rewrites in place
• Composer Multi-file Editing: Describe requirements in conversation, AI simultaneously modifies 20+ files and handles dependencies
• Agent Mode: Autonomously search codebase, execute terminal commands, write and run tests, long tasks can span hundreds of steps
• Codebase Indexing: Index the entire project, provide architecture-aware suggestions
• Grok 4.5 Native Integration (New): Launched on 2026/7/9, jointly trained model by SpaceX×Cursor, based on trillions of Cursor developer data, Token efficiency 4x that of Opus 4.8, directly callable within the editor
• Multi-model Switching: Besides self-developed Composer 2 and Grok 4.5, supports top models like Claude Opus 4.8, GPT-5.6, Gemini 3.1 Pro
• VS Code Ecosystem Compatibility: Can directly import VS Code extensions, themes, shortcuts, settings, zero migration cost

Use Cases

• Full-stack / senior developers coding 4+ hours daily
• Daily development of frontend / React / Next.js projects
• Rapid prototyping and MVP construction
• Multi-file refactoring and technical debt cleanup in large codebases
• Productivity tool for independent developers and startup teams

Pros

• Self-developed Composer 2 brings coding model price down to 1/10 of contemporary Opus flagship
• Industry-leading multi-file refactoring (can modify 20+ files simultaneously)
• Zero-cost migration from VS Code ecosystem
• Unlimited Composer usage in Auto mode, no credit deduction
• Extremely fast product iteration (Composer 1.5 → 2 → Cursor 3 UI restructured within same quarter)
• $50 billion valuation + top VC backing, product unlikely to disappear in short term

Pricing

In 2026, there are 4 tiers: **Free** ($0, 2-week Pro trial + 2000 completions + 50 slow requests), **Pro $20/month** (unlimited completions + 500 fast requests + unlimited slow), **Ultra $200/month** (priority features + maximum request volume), **Business $40/user/month** (team management, centralized billing, privacy mode, SSO). Composer 2 API prices: Standard $0.50/$2.50, Fast $1.50/$7.50 (cache read $0.20~$0.35). ⚠️ Since 2025-08, changed from "request limit" to "usage-based credit system"; unpredictable cost is currently the biggest complaint; can reduce credit consumption by setting Composer 2 as default in Auto mode.

Summary

Cursor is currently one of the **top 2 AI coding products** (the other being Claude Code), with almost no rival in daily coding / multi-file refactoring / rapid iteration scenarios, especially Composer 2 which brings top-tier coding capability price down to one-tenth of the contemporary Opus flagship. **Starting 2026/7/9, a new Grok 4.5 option is added** (jointly trained by SpaceX×Cursor, with overwhelming Token efficiency), further enriching the model matrix. However, it is not a no-brainer recommendation—credit-based pricing, reliance on third-party base models, and the catch-up of Claude Code are risks to be acknowledged. Selection advice: Choose Cursor for daily coding (can pair with Grok 4.5 for ultimate cost-effectiveness), Claude Code for complex autonomous tasks and deep reasoning, and Copilot for limited budgets. For the latest UI progress, see sub-entry "Cursor 3".

Version History

• Cursor launches Projects (beta), a coordinator agent directing thousands of subagents on large development tasks (2026-09-10): Cursor launched Projects in beta, rolling out to all users. You chat with a project's coordinator agent, which writes no code itself but delegates tasks to thousands of subagents running in parallel, so it is never blocked and always responsive. A project runs on its own cloud computer by default and keeps working when you close your laptop; when something needs local testing, the coordinator spins up a local agent. Each project maintains shared context synced across cloud and local machines, so research findings, codebase knowledge and user preferences recorded by one agent are reused by future agents. It also supports subscriptions: watching a Slack channel, running on a schedule, or following all your PRs to fix CI automatically. Internally, Cursor users who primarily use Projects merge about six times as many PRs, and new users merge 30% more.
• Cursor launches Self-Hosted Machines, letting cloud agents run on a company's own machines (2026-09-02): Cursor releases Self-Hosted Machines, migrating the tool execution of cloud agents to machines within the enterprise's own network. The agent loop, reasoning, and planning remain on Cursor's cloud, connected via outbound HTTPS from workers, and Cursor does not proactively connect into the enterprise network.
• OpenAI decides to cut off model access to Cursor over compliance risks following the SpaceX acquisition (2026-08-28): OpenAI has notified SpaceX that it will terminate the contract to provide OpenAI models to Cursor, with a proposed shutdown date of November 12, 2026, and has given the longest notice period allowed under the contract. OpenAI stated that this move stems from its inability to be confident that SpaceX will comply with the terms of service, citing precedents of Musk's companies previously violating contracts and xAI violating OpenAI's service terms. OpenAI expressed that it will make every effort to support affected developers during the transition.
• Cursor responds to OpenAI's plan to block its model access (2026-08-29): We regret to see that OpenAI has issued a statement announcing plans to block Cursor users from accessing OpenAI models within three months. OpenAI models carry approximately 5% of Cursor's user traffic, and we are communicating with the OpenAI team to resolve this matter. Cursor has been one of OpenAI's earliest users, and over the years we have worked closely with their team, consistently trusting their platform as neutral infrastructure for our business.
• Cloud Agents can now create projects from scratch without connecting GitHub or other third-party repositories, and save them to Cursor Origin (2026-08-27): Cloud Agents now supports creating projects from scratch without connecting to third-party repositories such as GitHub, and allows saving them to the Cursor Origin repository.
• Cursor posts a Changelog update enhancing Cloud Agents and Cursor Harness: cloud agents can respond automatically (2026-08-19): Cursor released a Changelog update, enhancing Cloud Agents and Cursor Harness capabilities: cloud agents can automatically respond to events and pick up work, persistently hold goals until completion, and support long-running sessions; a new Subscriptions feature is added, enabling monitoring of PRs, Slack threads, or running scheduled tasks.
• Cursor launches Origin, a code hosting service positioned as a GitHub alternative (2026-08-17): Starting today, Cursor is opening the early beta of Origin code hosting to all paid plan users, offering repositories, pull requests, code browsing, and GitHub sync features. Users can create repositories prefixed with cursor.com/codebase/ or sync GitHub repositories to Origin, with bidirectional sync of comments and reviews. Integrations with Vercel, Depot, and Buildkite are now available, with agent features coming soon.
• Cursor launches builds: cloud agents now start up to 3x faster (2026-08-13): Cursor launches the builds feature, which continuously prepares ready-to-use copies of development environments in the background, allowing cloud agents to start without building from scratch and improving response speed by up to 3 times. Internal environments start 10 times faster, and the first token is generated 3 times faster; agents always start from the most recent successful build, so dependency updates or installation script errors won't affect runtime. Starting August 17, builds will be enabled by default for all environments at no additional cost.
• Cursor releases Cursor Router, an intelligent model routing system (2026-07-22): Cursor has launched Cursor Router, which automatically assigns each coding request to the most suitable model. Online A/B testing shows that the Auto Intelligence mode achieves user satisfaction close to Fable while reducing costs by approximately 60%; the Auto Balance mode surpasses Opus 4.8 in satisfaction, with a cost reduction of about 36%.
• Cursor Router Intelligent Model Routing (2026-07-22): Automatically assigns coding requests to the most suitable model; Auto Intelligence reduces cost by ~60%, Auto Balance satisfaction exceeds Opus 4.8 with 36% cost reduction
• Grok 4.5 Integration (2026-07-09): Jointly trained model by SpaceX×Cursor launched and natively integrated into the editor, 1.5T parameter V9 base model, Token consumption for equivalent tasks only 1/4 of Opus 4.8