Japanese startup Sakana launches Fugu multi model API beating rivals
Sakana, a Japanese AI startup, has launched Fugu, a new system that handles multiple models through one API for tough tasks.
Its advanced version, Fugu Ultra, just beat Anthropic's Claude 5 on coding tests like LiveCodeBench (scoring 93.2 vs. Fable's 89.8) and surpassed Mythos Preview on graduate-level science benchmarks, showing it's pretty sharp at solving dynamic problems.
Fugu Ultra tops GPQA-D test
Fugu Ultra also topped the GPQA-D (Diamond) graduate-level science test with a score of 95.5, edging past Mythos Preview's 94.6.
Sakana says Fugu even outperformed Google's Gemini 3.1 Pro and OpenAI's GPT-5.5 in things like Japanese handwriting analysis, financial forecasting, and Rubik's Cube solving.
Meanwhile, Anthropic pulled Claude Fable 5 from the market after US officials worried it could be misused for bioweapons or cyberattacks, so now models like Fugu are leading the pack in performance benchmarks.