Step launches Step5Preview, a 600B sparse MoE model optimized for real-world Agentic tasks. Activating only 27B parameters, it supports 1M token context and multimodal inputs. Scoring top-three globally, it costs one-eighth of Claude Opus5 while matching performance in coding, long-horizon reasoning, and finance. Key benchmarks include surpassing Opus5 in GPU kernel optimization and demonstrating robust multi-step agent workflows. Full open-source weights release October 15, positioning Step5Preview as a high-efficiency flagship for daily agent use.