A Japanese AI lab has reportedly introduced a system called Sakana Fugu, designed to compete with some of today’s most capable frontier models. The project comes from a team that includes a co-author of the original Transformer paper, the foundational breakthrough behind modern large language models. What makes Fugu stand out isn’t just benchmark performance, but its architecture. Instead of relying on a single large model, Fugu functions as a coordinator across multiple frontier models. When a user submits a request, the system dynamically decides which models are best suited for different parts of the task, routes subtasks accordingly, and then merges their outputs into a single, unified response. In effect, it treats AI models not as competitors, but as specialized components within a larger system. Reported SWE-Bench Pro results suggest Fugu Ultra achieved a score of 73.7, compared to 69.2 for Claude Opus 4.8 and 58.6 for GPT-5.5, highlighting its potential strength in complex...