🚨JAPANESE AI STARTUP JUST MATCHED CLAUDE FABLE 5 AND MYTHOS PERFORMANCE.
Japanese AI lab just launched Fugu, a model trained to command other models.
Sakana AI, a Tokyo-based AI startup, was co-founded by researchers including one of the authors of the original Transformer paper, the blueprint that basically every AI model today is built on.
Sakana Fugu is a multi-agent AI system that feels like a single model, automatically orchestrating specialized models behind one API endpoint.
Fugu Ultra is claimed to match Anthropic's Fable 5 & Mythos Preview on the hardest engineering/science/reasoning benchmarks — and to beat Gemini 3.1 Pro, Opus 4.8 & GPT-5.5 on tasks like AutoResearch, mechanical design & financial forecasting.
Everyone's racing to scale ONE giant model but Fugu flips it: an LLM that orchestrates a pool of the world's best models (choosing who does what, delegating, verifying and merging it into one answer), even calling itself recursively.
Think of it as a conductor. You send one request to one API, and Fugu figures out the move: answer it solo, or assemble a squad of expert models and run the whole thing for you.
Sakana AI launched two models:
Fugu - fast, low-latency, everyday coding/chat
Fugu Ultra - max quality on hard, multi-step problems (AI research, paper reproduction, cybersecurity, patent search)
Export controls are hitting frontier models now, but because Fugu's agent pool is swappable, if a provider restricts access, it routes around it.
Both Fugu and Fugu Ultra can be accessed through a unified API, with subscription plans for everyday users and pay-as-you-go pricing designed for high-volume and enterprise workloads.