SWE-1: Our First Frontier Models

The newly launched SWE-1 family of models aims to optimize the entire software engineering process beyond mere coding tasks, featuring three distinct models: SWE-1, SWE-1-lite, and SWE-1-mini. Each model is tailored for specific use cases, with SWE-1 offering advanced tool-call reasoning at a competitive cost, SWE-1-lite providing enhanced quality as an unlimited free option, and SWE-1-mini designed for fast, passive predictive tasks. These models represent a significant advancement in coding capabilities, transitioning from simple autocomplete functions to developing complete applications autonomously. Despite not reaching the absolute frontier, SWE-1 competes closely with leading foundation models, outperforming non-frontier and open-weight alternatives through both offline evaluations and blind production experiments. With a focus on user interaction and incremental improvements in flow awareness, the initiative underscores a commitment to surpassing current frontier models by investing in enhanced model performance and cost-effective solutions for software engineering tasks.