Cursor just Became UNSTOPPABLE

TL;DR

• 1

  Cursor is training an Opus-class model from scratch: Wes says the new model uses 10x to 20x more compute than Cursor's earlier Composer models and is meant to compete directly with Claude Opus and GPT-class coding systems.

• 2

  The SpaceX-xAI-Cursor tie-up solves each side's main weakness: SpaceX and xAI bring Colossus-scale compute, while Cursor brings a huge coding user base, workflow data, and a product that was previously too dependent on OpenAI and Anthropic APIs.

• 3

  Cursor's edge may be reinforcement learning, not just raw scale: Wes highlights Cursor's targeted textual feedback method, where a teacher model inserts local hints at the exact step an agent made a bad tool call, wrong edit, or style mistake.

• 4

  Composer already proved the recipe can work cheaply: Cursor reportedly started from Kimi K2.5, added extra training and RL, and produced Composer 2 and 2.5 as strong coding models with much lower inference cost than frontier Western models.

• 5

  GitHub's outages created an opening for Origin: Cursor's new Origin product is framed as Git for AI agents, built for agentic workflows after GitHub was hit by what Wes calls an unexpected tidal wave of AI activity and reliability problems.

• 6

  This is also a test of whether a so-called wrapper can become a lab: Wes points out that Cursor began on Microsoft's open-source VS Code, depended on frontier APIs, and did not own compute early on, yet may now be snowballing into one of the strongest model builders in coding.