SpaceX acquired Anysphere for $60 billion in August 2026. This acquisition follows a period of extreme growth where the company reached $2 billion in annualized revenue in February 2026 and climbed to $4 billion annualized by mid-2026. The company currently manages 1 million paying customers and 50,000 enterprise teams. Enterprise contracts now account for 75% of total revenue. The rapid scaling of the business exceeds previous benchmarks for software companies like Slack or Zoom.
The transition from individual developers to large engineering organizations defines the current market. Cursor is a fork of Microsoft Visual Studio Code. This inheritance allows the editor to include language servers and existing extensions. The product uses a combination of proprietary models and third-party models to complete tasks. The company reached $100 million in annualized revenue in January 2025 and $1 billion by November 2025.
The market for AI coding tools is expanding rapidly. AI coding tools generated $12.8 billion in revenue in 2026. More than half of all code on GitHub is now AI-generated or AI-assisted. Ninety percent of developers use at least one AI tool at work.
The two-pool credit system and usage costs
Cursor uses a two-pool credit system for its paid plans. One pool provides usage for Cursor Models like Grok 4.7, Grok 4.6, and Composer 2.5. The other pool covers third-party models from providers such as Anthropic and OpenAI. Third-party model requests incur a Cursor Token Rate of $0.25 per million tokens. This rate applies to input tokens, output tokens, and cached tokens on eligible requests.
The pricing structure changed several times after June 2025. The company moved away from a flat request-based model to a usage-based system to maintain gross margins. The Pro plan costs $20 per month. The Team plan costs $40 per user per month and includes features like centralized billing and SAML/OIDC SSO. The Enterprise plan is available via custom pricing for those who need advanced security controls or pooled usage.
| Model Name | Input Price (per 1M) | Cache Write (per 1M) | Cache Read (per 1M) | Output Price (per 1M) |
|---|---|---|---|---|
| Claude 3.5 Sonnet | $2 | $2.5 | $0.2 | $10 |
| Claude Opus 5.5 | $4 | $5 | $0.2 | $20 |
| GPT-5.6 Luna | $0.2 | $0.25 | $0.02 | $1.2 |
| Grok 4.7 | $2 | $0 | $0.5 | $6 |
| Gemini 3.1 Pro | $2 | $0 | $0.2 | $12 |
You should check your usage dashboard frequently to avoid unexpected costs. Heavy use of Agent mode draws down the credit pool faster than autocomplete usage. When a user consumes all included third-party usage, Cursor switches them to the Cursor Models pool. If they continue to use third-party models, they use on-demand usage which bills at model rates.
Context window and monorepo reliability limits
The ability of the editor to understand a codebase depends on its indexing capabilities. The @codebase feature works reliably for repositories with fewer than 200k lines. Users report inconsistent results when they use this feature on repositories between 200k and 500k lines. For monorepos above 500k lines, the indexing becomes unreliable for code navigation and change planning. The tool is unreliable for large monorepos.
Large tasks often hit the context window limits of the underlying models. A task that requires reading 30 files simultaneously can exhaust the available tokens. This leads to a loss of coherence in the AI response. Users must break large tasks into smaller sub-tasks to maintain accuracy.
Context management is a primary limitation for the editor. When a task spans many files, the system loses track of earlier instructions. This creates errors in complex features. The tool is optimized for application-scale codebases rather than platform-scale monorepos.
The distinction between Plan, Agent, and Composer modes
Cursor provides three distinct ways to interact with AI. The Chat mode uses Cmd+L to allow for questions about code. Inline Editing uses Cmd+K to modify selected blocks of code. Composer uses Cmd+I to allow for multi-file edits.
Composer acts as a director for multi-file tasks. The user selects files and describes the required changes. The system shows a diff for each file. The user must then review and accept or reject these changes. Plan mode allows the AI to create a detailed implementation plan before it writes any code. This mode is designed to prevent the AI from wandering during complex tasks. Agent mode allows the AI to work autonomously. It can create files, edit code, and execute terminal commands.
I find that Plan mode is the best way to manage non-trivial work. Agent mode is better for quick changes or tasks the user has performed many times before. The three modes serve different stages of the development lifecycle.
The danger of blind acceptance and security risks
AI assistants produce code that can include security vulnerabilities. They can produce working code that skips access control or leaves row-level security unconfigured. They can also produce an API that does not exist. When a developer accepts a fifty-line diff without reading every single line of code, they risk maintaining a codebase full of logic errors and security vulnerabilities that they do not fully understand or can no longer control.
Reviewing the diff is a necessary part of the workflow. The tool generates a diff view before any changes are applied. The user must read this diff to catch mistakes. Relying on the AI to handle sensitive logic like authentication or payment flows without a line-by-line pass is a mistake.
The productivity gains from AI are visible in pull request velocity. However, these gains disappear if the developer must spend hours fixing AI hallucinations. Developers should always verify AI suggestions against real documentation.
Technical failures of Tab autocompletion
The Tab autocompletion feature uses a custom fine-tuned model for speed. This feature predicts multi-line completions and suggests auto-imports. Some users report that Tab autocompletion stops working after an update. This often happens due to extension conflicts or network issues.
Users can troubleshoot this by checking the status bar icon. The settings menu allows users to ensure that Cursor Tab is enabled. If the feature is paused, the user can unpause it in the settings. Another fix involves disabling HTTP/2 in the application settings. If extensions cause the problem, the user can run the command cursor --disable-extensions from the command line.
The autocomplete performance is fast. It achieves latency between 150ms and 250ms on modern hardware. This is much faster than the 400ms to 800ms latency seen in GitHub Copilot.
The reality for non-technical founders
Non-technical founders often expect Cursor to build a finished, production-ready product. This is an incorrect assumption. The tool is a code editor. It requires the user to read code and evaluate if a change is correct.
The tool cannot handle the remaining 30 to 40 percent of engineering work. This work includes setting up multi-role authentication and managing data migrations. It includes handling retries, webhooks, and timeouts. The tool produces code, but it does not produce a complete business system.
Founders who cannot debug code will find that they spend all their time reviewing incorrect changes. The tool makes a competent developer faster, but it does not replace the need for engineering judgment.
Comparison of professional workflows
The choice between tools depends on the environment and the workflow. GitHub Copilot is an extension that runs in VS Code, JetBrains, and Neovim. It is a good choice for teams that want to stay in their existing editor. Cursor is a fork of VS Code. It is a good choice for developers who want deep integration and better autocomplete speed.
Windsurf provides an agentic workflow engine called Cascade. It offers about 80 percent of the capability of Cursor at 75 percent of the price. Claude Code is a terminal-based agent. It is a good choice for developers who live in the shell.
| Tool | Shape | Best for |
|---|---|---|
| Cursor | AI-native editor | Developers using VS Code who want the best autocomplete |
| GitHub Copilot | AI extension | Teams using various editors like JetBrains or Neovim |
| Windsurf | AI-native editor | Developers who prefer an agentic workflow |
| Claude Code | Terminal agent | Developers who work primarily in the shell |
Will the current pricing structure hold once more models reach parity?
