Anysphere teases a massive, from-scratch Cursor model

If you thought Cursor was content just being the best AI code editor on your laptop, think again.

At its Compile event, Cursor maker Anysphere pulled back the curtain on three massive projects. Leading the charge is a brand-new, proprietary foundation model trained completely from scratch. The startup, which was recently acquired by SpaceX, confirmed the model is currently in training and should ship within the next few weeks.

Building an Opus rival from scratch

Cursor co-founder Michael revealed that this is the team's first fully self-trained model. Unlike earlier Composer iterations that relied on open-source bases or Kimi, this one is built from the ground up.

It matches the sheer size of Claude Opus and GPT-4.5, burning through 10 to 20 times more compute than Cursor's previous models. Interestingly, Anysphere says it is not just built for code, but designed with general intelligence to handle broad tasks across the board.

Origin: Git built for thousands of AI agents

Training a model was only part of the show. Cursor also announced Origin, a reimagined Git platform designed for seamless collaboration between human engineers and autonomous agents.

Co-founder Tomas, who joined through the Graphite acquisition, explained that Origin runs on a fresh Git architecture layered over cloud providers. During load tests, the system handled thousands of agents reading and writing to a single repository simultaneously. Origin automatically fixes merge conflicts, patches failed CI runs, and manages PR comments. It is already live internally and with select partners, with a broad rollout slated for fall.

To top it off, Cursor launched Cursor Mobile in beta for iOS. The app lets developers babysit their agents remotely, inspect agent-generated screenshots, resolve stuck prompts, and control locally running workers right from their phone.

Why it matters

Cursor is clearly tired of just wrapping other companies' LLMs. By leveraging SpaceX's resources to build a full-scale foundation model, the team is moving to control its entire tech stack from the brain to the editor.

For developers, Origin might actually be the most disruptive piece here. Standard Git workflows break down quickly when multiple agents start spamming pull requests and generating merge conflicts. If Origin can reliably coordinate thousands of agent commits and auto-heal broken builds, developer productivity could leap far beyond what a simple in-editor copilot offers.

That said, training a frontier-tier model from scratch is notoriously risky and expensive. Anysphere will have to prove its new weights can actually outperform Claude and OpenAI on real-world reasoning. But by combining proprietary intelligence, an agent-first version control platform, and mobile orchestration, Cursor looks like it is building the default operating system for AI software development.