Is Cursor Losing Its Spotlight Despite $500M ARR?

Context and market shift

Cursor 2.5 was released with a plugin marketplace, sandbox network‑access controls, and asynchronous sub‑agents, representing a high‑content major version. Despite these upgrades, community discussion volume decreased noticeably.

Financial data shows Anysphere’s ARR exceeded $500 million, a $9 billion valuation, and 30 000 NVIDIA engineers using Cursor, reporting a three‑fold increase in code output (February data).

The product’s core premise is human‑AI collaboration: developers write code while AI provides completions, refactoring suggestions, and debugging assistance, positioning AI as a co‑pilot.

Industry narrative shift

From mid‑2025 the narrative moved toward Anthropic’s Claude Code, a terminal‑only tool that clones repositories and submits pull requests without an editor. Claude Code’s annualized run‑rate grew from $10 billion to $25 billion by February 2026, roughly five times Cursor’s revenue.

OpenAI’s Codex App (released February 2024) follows the same logic, letting a user command multiple AI agents in parallel; the editor becomes optional.

This shift makes the “AI‑native IDE” story less novel, prompting Cursor to emphasize its emerging autonomous‑agent capabilities hinted at in the 2.0 roadmap.

Dependency on third‑party models

Cursor relies on Anthropic’s Claude and OpenAI’s GPT for core capabilities. Both providers now ship end‑user coding products (Claude Code, Claude Cowork, Claude Safety Agent; OpenAI Codex App, announced as free), creating a “sandwich” risk where Cursor sits between model developers and downstream users.

Similar model‑dependency challenges affect Codeium (Windsurf) and ByteDance’s Trae.

Competitive pressures

Open‑source alternative OpenClaw : gained 150 k GitHub stars in its first week, supports arbitrary models, runs locally at zero cost, offering a privacy‑focused AI‑agent platform.

Pricing competition : ByteDance’s Trae offers a free‑month trial; Windsurf’s pricing is roughly half of Cursor’s $20 / month Pro tier and includes a free tier.

IDE vendors : JetBrains bundles AI assistants across its suite; VS Code ships with GitHub Copilot and now integrates Claude Code natively, reducing the incentive to switch editors.

Cursor 2.5 Feature Highlights

Plugin Marketplace

The marketplace bundles skills, sub‑agents, MCP servers, hooks, and rules for one‑click installation. Early partners include Amplitude, AWS, Figma, Linear, and Stripe. Example: the Figma plugin fetches design files and generates code directly in Cursor; the Stripe plugin creates payment integration with a single command.

Browse https://cursor.com/marketplace or install plugins via the /add-plugin command.

Sandbox Access Controls

Admins can whitelist specific domains, combine user‑specified domains with Cursor’s defaults, or allow unrestricted access. Enterprise administrators can enforce organization‑wide allow/deny lists through a management console, addressing security when AI agents access internal code or production environments.

Asynchronous Sub‑agents

Building on the synchronous sub‑agents introduced in Cursor 2.4, the new async version lets parent agents continue work while sub‑agents run in the background. Sub‑agents can spawn their own sub‑agents, forming a tree‑structured collaboration that breaks large tasks into parallel subtasks.

This enables AI to coordinate multiple agents without leaving the editor, responding to the trend of autonomous coding.

Other Improvements

Conversation Search : search historical dialogues.

Plan Mode Cloud Build : execute complex plans in the cloud.

Pre‑configured Commands : git clone, npm install, pip install, etc., are available out of the box.

Inline Diff Switch : toggle code‑diff display modes.

Performance optimizations for sub‑agents: lower latency and smoother streaming feedback.

Strategic implications

Cursor’s 2.5 release demonstrates a shift toward platformization and autonomous‑agent features to mitigate narrative loss and model‑dependency risk. The product’s future success will depend on whether IDEs can remain the primary entry point in an AI‑agent workflow where natural‑language interfaces increasingly replace traditional editors.