What Happened

Cursor 3 has been released with a significant architectural shift: the AI code editor is ditching its VS Code foundation and launching its own proprietary frontier model. The new model has posted strong results on several coding benchmarks, though the methodology behind those benchmarks has drawn skepticism from developers who question whether the numbers reflect real-world performance.

Why It Matters

For indie developers and SMEs, this move has two direct consequences. First, forking away from VS Code means Cursor is now building and maintaining its own editor core, which increases development velocity for Cursor-specific features but also increases the risk of ecosystem fragmentation. Extensions, keybindings, and workflows built around VS Code compatibility may break over time.

  • Proprietary model means no transparency on training data, context window behavior, or pricing trajectory
  • Benchmark dominance without reproducible methodology is a red flag for teams making procurement decisions
  • Switching costs rise as Cursor diverges further from the VS Code API surface

Asia-Pacific Angle

Chinese and Southeast Asian development teams evaluating AI coding tools should note that Cursor's proprietary model raises data residency concerns. Unlike open-weight alternatives such as Qwen-Coder or DeepSeek-Coder, which can be self-hosted within a local cloud region, Cursor's model sends code to US-based infrastructure. For teams in regulated industries or those building for government clients in markets like Singapore, Indonesia, or mainland China, this is a compliance consideration worth documenting before adoption. Domestic alternatives including Tongyi Lingma (Alibaba) and Baidu Comate are worth benchmarking against Cursor 3 on the same internal codebases before committing.

Action Item This Week

Run a controlled comparison: take three real tasks from your current sprint, complete each one using Cursor 3, your current editor with GitHub Copilot, and one open-source alternative like Continue.dev with a self-hosted Qwen-Coder model. Record time-to-completion and error rate. Use that internal benchmark, not vendor-published numbers, to guide your tooling decision.