How Sakana's Fugu beats Anthropic's Fable without using it
sakana ai launched fugu, a multi-llm orchestrator matching anthropic's fable 5 and mythos preview on benchmarks. this signals a shift toward composite architectures that mitigate vendor lock-in and regulatory risk.
sakana ai introduced fugu, a system that behaves like a single model but internally routes tasks to a swappable pool of specialized llms via an openai-compatible api. benchmarks show fugu ultra scoring 73.7 on swe bench pro and matching anthropic's fable 5 on reasoning tasks, outperforming individual base models like opus 4.8 and gemini 3.1 pro. the architecture addresses vendor lock-in, allowing organizations to swap agents if access is revoked by export controls. early adopters are applying the ultra variant to cybersecurity, patent search, and reproducing scientific papers.