Japan's AI startup SakanaAI has launched Fugu, an AI system that does not rely on a single model but orchestrates multiple AI models through a single API to handle complex tasks, NDTV reported on Monday.

SakanaAI said Fugu Ultra delivered comparable performance to Anthropic Fable5 and Mythos Preview on core engineering, science and reasoning benchmarks, and outperformed Fable5 in some categories. Published benchmark results showed Fugu Ultra scored 93.2 and Fugu scored 92.9 on LiveCodeBench, which tests coding performance, beating Fable5's 89.8. In GPQA-D, which covers graduate-level problems in biology, physics and chemistry, Fugu Ultra and Fugu scored 95.5, topping the previous Mythos Preview model's 94.6.

SakanaAI released two versions, Fugu and Fugu Ultra. Fugu is aimed at everyday tasks such as coding and chat, while Fugu Ultra focuses on more complex work including AI research, reproducing academic papers, cybersecurity analysis and patent searches.

SakanaAI said its own tests showed the Fugu models outperformed Google Gemini 3.1 Pro, OpenAI GPT-5.5 and Anthropic Opus 4.8 in areas including automated research, machine design, Japanese handwriting analysis, chess, solving a Rubik's Cube and financial time-series forecasting.

Tokyo-based Sakana AI was co-founded in 2023 by Llion Jones (리온 존스), a co-author of Google's 2017 paper "Attention Is All You Need", and David Ha (데이비드 하), a former head of research at Stability AI.