Japan’s Sakana Launches Fugu and Fugu Ultra Models

Japanese AI startup Sakana launched a new system called Fugu that outperformed Anthropic's Claude Fable 5 on key coding benchmarks, a notable result given that Fable 5 was pulled back by the U.S. government just three days after its release.

Fugu is not a single model. It coordinates multiple AI models through one API to solve complex tasks, an approach that reflects growing interest in system-level intelligence rather than monolithic models. Sakana released two versions. Fugu handles coding, chat, and everyday tasks, while Fugu Ultra is aimed at AI research, cybersecurity analysis, patent investigations, and paper reproduction.

On LiveCodeBench, an open-source benchmark that tests coding performance on regularly refreshed problems, Fugu Ultra scored 93.2 and the base Fugu scored 92.9, both ahead of Fable 5's 89.8. On GPQA Diamond, a set of 198 graduate-level science questions, both Fugu variants scored 95.5, beating the earlier Mythos Preview model's 94.6. Sakana also claimed its models surpassed Google's Gemini 3.1 Pro, OpenAI GPT-5.5, and Anthropic's Opus 4.8 on tasks ranging from automated research and mechanical design to financial forecasting.

Anthropic had released Fable 5 as a guarded version of its underlying Mythos model, which the company deemed too powerful for mass release. Mythos was found to identify vulnerabilities in every major operating system and browser tested, some decades old. The U.S. government then asked Anthropic to revoke foreign access, cutting Fable 5's availability after only three days.

Sakana was founded in 2023 by Llion Jones, a co-author of Google's 2017 "Attention Is All You Need" paper, and David Ha, former head of research at Stability AI. The startup has positioned itself around nature-inspired algorithms and multi-agent AI design, and the Fugu launch indicates that smaller, coordinated systems can challenge the largest closed models, especially when those models face regulatory constraints.