Cursor AI Launches Composer 2.5, Partners with SpaceXAI for Advanced Model Training

Cursor AI recently announced the release of Composer 2.5, marking a substantial advancement in its AI-assisted coding capabilities. This latest iteration of their developer-focused platform showcases impressive performance metrics, reportedly achieving results comparable to top-tier models such as Claude Opus 4.7, while offering a significantly more cost-effective solution, up to ten times cheaper for complex workloads. The core improvements in Composer 2.5 stem from sophisticated training methodologies. Cursor AI has implemented scaled reinforcement learning (RL) with targeted textual feedback during model rollouts. This approach allows the AI to pinpoint and correct errors within long operational trajectories, leading to a notable increase in reliability and fewer false starts during collaborative coding sessions. Furthermore, the model was trained on 25 times more synthetic tasks, meticulously grounded in real-world codebases, including scenarios like feature deletion and reimplementation with test-based rewards. This rigorous training has resulted in improved effort calibration and reduced hallucinations when utilizing tools. Beyond the technical enhancements, a major strategic development for Cursor AI is its newly forged partnership with SpaceXAI, a collaboration involving xAI and SpaceX. This alliance is set to enable Cursor to train a considerably larger foundational model from the ground up. The initiative will leverage SpaceX's formidable Colossus 2 data center, which boasts an equivalent of one million H100 GPUs, providing immense computational power for this ambitious undertaking. This partnership underscores a strategic move to push the boundaries of AI in coding and knowledge work, potentially transforming how developers interact with and create software. Composer 2.5 is immediately available within the Cursor platform, with users receiving doubled usage allowances for the initial week. The pricing structure for the standard model is set at $0.50 per million input tokens and $2.50 per million output tokens, with a faster, albeit slightly more expensive, variant also available, both remaining more economical than competing fast-tier models. The underlying technology for Composer 2.5 builds upon the open-weight Moonshot AI Kimi K2.5 base, with Cursor contributing approximately 85% of the post-training compute. This release and partnership position Cursor AI as a significant player in the evolving landscape of AI-driven developer tools, aiming to enhance productivity and streamline complex software development workflows through advanced AI models and strategic collaborations.