Cursor Unveils In-House AI Model, Git Platform and Mobile App
Cursor unveiled its first from-scratch AI model, an agent-first Git platform called Origin and an iOS app, days after SpaceX’s $60 billion acquisition of its parent.
Cursor used its first Compile conference in San Francisco to unveil three products at once: its first AI model trained entirely in-house, a Git hosting platform built for AI agents, and a mobile app.
The announcements signal a push by the AI coding company to control more of the software stack, and they landed the same day SpaceX confirmed a $60 billion all-stock deal to acquire Cursor’s parent, Anysphere, placing it under the xAI unit. Cursor, founded in 2022 by four MIT graduates, has grown fast, reaching more than $3 billion in annual recurring revenue by May, up from $100 million in early 2025.
The centerpiece is a frontier model of more than 1.5 trillion parameters, a scale CEO Michael Truell said is comparable to Anthropic’s Opus and OpenAI’s GPT series. Unlike Cursor’s earlier Composer models, which were built on open-source bases such as Kimi K2.5, this one is being trained from scratch on xAI’s Colossus supercomputer in Memphis, using more than 100,000 GPUs and 10 to 20 times more compute than any prior Cursor model. Truell positioned it as general-purpose, “intelligent beyond coding,” and said training is underway with a release expected within weeks. Cursor has not published benchmarks.
The second product, Origin, is a Git platform built on the assumption that AI agents, not humans, are the primary users. Built on technology from Graphite, the code-review startup Cursor bought in December, and demoed by Graphite co-founder Tomas Reimers, Origin uses an AI engine to resolve merge conflicts, fix failed tests and handle comments at high concurrency.
Cursor showed demo figures of about 22 commits per second and 296,000 clones an hour in a single repository, numbers it has not independently verified. It is running internally and with select partners, with broad availability planned for the fall. The third launch, Cursor Mobile, is an iOS beta that lets developers supervise agents from a phone, unblocking stuck tasks and reviewing agent-generated diffs and screenshots.
The Vertical Integration Bet
Together the products describe a company trying to own the full development pipeline, from the editor where agents write code to the model that powers them and the host where the code is stored.
For years Cursor’s code lived on Microsoft’s GitHub and its AI ran through models from Anthropic and OpenAI. Owning all three layers creates a data flywheel rivals cannot easily copy and targets a real strain: GitHub now processes hundreds of millions of AI-agent commits a week on infrastructure designed for humans.
The risk is execution. Building a genuinely agent-native version control system is a multi-year effort, the demo metrics are unverified, Origin has no pricing yet, and GitHub’s network of more than 100 million developers is hard to dislodge.
The SpaceX Factor
The SpaceX acquisition, expected to close in the third quarter pending regulatory approval, reshapes the stakes. It gives Cursor access to xAI’s Colossus cluster, the same hardware now training its model, but it also raises pointed questions for developers. Cursor built its reputation partly on model neutrality, letting users route work to Claude, GPT or its own models.
After the deal closes under xAI, whether Cursor keeps supporting rival models is an open question, though no changes have been announced and it still offers Claude and GPT today. Community reaction also flagged unresolved data and privacy terms, with users wary that code hosted on Origin could be used to train models. For now, the vision is clear and the guarantees are not.
Anthropic is hiring across Australia and Japan to expand its AI data center capacity in Asia-Pacific, as soaring demand for Claude strains its infrastructure.
