You know Cursor as the AI coding editor that took over developer desktops. A VS Code fork — a modified copy of Microsoft's free code editor — that plugged in models from Anthropic and OpenAI and charged $20 a month for the privilege. A smart wrapper, essentially.
The problem with wrappers: they exist at the mercy of whoever makes the thing being wrapped. Anthropic can launch Claude Code. OpenAI can ship Codex. Google can build Gemini into every IDE (integrated development environment — the software developers write code in). If your entire product is someone else's model behind your UI, you are one strategic decision away from irrelevance.
Cursor just made its counter-move. On March 19, the company shipped Composer 2 — its own frontier coding model built for multi-file edits, refactoring, and long-running agentic tasks. Six days later, it launched self-hosted cloud agents — letting enterprises run Cursor's AI agents entirely inside their own networks, with code, secrets, and build artifacts never leaving company infrastructure.
Composer 2 scores 61.7 on Terminal-Bench 2.0 — a standardized test for AI coding agents maintained by the Laude Institute — beating Claude Opus 4.6 at 58.0. It costs $0.50 per million input tokens versus Opus 4.6's $5. Ten times cheaper for better coding performance, according to Cursor's benchmarks. GPT-5.4 still leads at 75.1, so Cursor is not claiming the crown outright — but it no longer needs to rent someone else's. The model supports 200,000-token prompts (roughly a 500-page book of code) and handles tasks requiring hundreds of sequential actions.
The self-hosted agents solve a different problem. Regulated enterprises — banks, healthcare, defense — cannot let source code leave their network. Cursor's architecture: a worker process connects outbound via HTTPS to Cursor's cloud for inference and planning, then executes all tool calls locally on your machines. No inbound ports, no VPN tunnels. Your code stays home. Brex, Notion, and Money Forward already run it in production.
These moves land on top of an already absurd March. The company hit $2 billion in ARR (annualized recurring revenue — what a subscription business would make if current sales held for a full year), doubling from $1 billion in 90 days. It expanded into JetBrains IDEs — IntelliJ, PyCharm, WebStorm — reaching 16 million developers who refused to leave their editor. Bloomberg reported talks for a new funding round at $50 billion, nearly double the $29.3 billion November valuation. Enterprise customers generate 60% of revenue. Coinbase, eBay, Datadog, Sentry — and yes, OpenAI — all pay for Cursor seats.
The tax is real, though. Cursor confirmed a code reversion bug this month — the agent silently undid user changes when contexts conflicted. Three root causes identified: Agent Review tab interference, cloud sync racing local saves, and format-on-save collisions after AI edits. The official workaround: close a core feature to prevent data loss. That is not a fix — that is a band-aid on an arterial wound. And Composer 2, for all its efficiency gains, still trails GPT-5.4 by a wide margin on the hardest benchmarks. When 67% of the Fortune 500 depends on a single coding tool, every bug and every model limitation propagates across the industry.
The pattern matters more than any single announcement. Cursor started as a UI layer on someone else's models. Now it owns a frontier model, a cloud agent runtime, integrations across both VS Code and JetBrains, and enterprise contracts that survive downturns. It is building a vertical stack — models, runtime, distribution — in a market where most competitors control only one piece.
Fourteen months ago, Cursor was a promising editor with a $100 million run rate. Today it trains its own models, runs agents on your infrastructure, and targets a $50 billion valuation. The wrapper grew teeth.
