Developers often face a tradeoff between speed and intelligence in AI coding. On October 16, a new model named SWE-1.5 was released, designed to optimize both aspects. It reconfigures the entire AI stack—model, inference, and agent harness—into a unified system. The development process incorporated end-to-end reinforcement learning on real tasks, continuous model iteration, and improvements to tools and systems. A focus on real-world feedback helped fine-tune the model, which was tested through multiple beta versions before its release. SWE-1.5 aims to overcome shortcomings in existing coding environments. Many labs rely on limited task distributions and narrow evaluation metrics that often incentivize low-quality code. To counter this, the team created a comprehensive dataset reflecting a wider range of real-world tasks and established rigorous grading mechanisms, including classical tests and agentic grading. This approach helps ensure a more robust evaluation process, reducing false positives and enhancing overall model performance. The new model is trained on advanced GB200 NVL72 chips, marking it as one of the first public models to utilize this hardware generation. The team implemented reinforcement learning with a focus on co-optimizing the model and harness, allowing for continuous adjustments based on real-world performance. SWE-1.5 shows significant improvements in coding benchmarks, particularly in speed, allowing tasks that previously took around 20 seconds to be completed in under 5 seconds. This model is now actively used by engineers for various applications, including navigating large codebases and building full-stack applications.