Sakana Fugu: One Model to Command Them All

Sakana AI has a new product called Fugu, and it's less a single model than a dispatcher that decides which model should actually do the work. You send one request to one API, and Fugu figures out whether to answer directly or assemble a crew of expert models — sometimes including copies of itself, called recursively — to handle something harder. The company frames this as collective intelligence rather than brute-force scale, and it's betting that knowing which model to use matters more than building one giant model to rule them all.

The launch comes in two flavors. Plain Fugu is built for speed and everyday use — coding, code review, chatbots — and even lets teams exclude specific agents from its pool if they've got privacy or compliance concerns. Fugu Ultra is the heavier option, tuned for messy, multi-step problems where accuracy matters more than latency. Early users reportedly leaned on Ultra for things like AI research, reproducing academic papers, cybersecurity analysis, and digging through literature and patents. On benchmarks, Sakana says Fugu Ultra stands shoulder-to-shoulder with Anthropic's Fable 5 and Mythos Preview across tough engineering, science, and reasoning tests.

The timing of that comparison isn't accidental. Sakana points directly at export controls recently imposed on Anthropic's Fable and Mythos models as proof that relying on one vendor's API for critical work is a real vulnerability, not a hypothetical one. Access, the company argues, can vanish overnight because of shifting regulations. Fugu's pitch is that because its underlying agent pool is swappable, it can route around a disruption like that instead of going dark with it.

Under the hood, this isn't just clever routing logic bolted onto existing models. Fugu is itself trained to know when to delegate, how agents should talk to each other, and how to merge their output into one coherent answer — work that builds on Sakana's own research, including two ICLR 2026 papers, Trinity and the Conductor. In testing across tasks like AutoResearch, Rubik's Cube solving, mechanical design, Japanese handwriting analysis, one-shot chess, and financial time-series prediction, Sakana says Fugu models came out ahead of frontier options including Gemini 3.1 Pro, Opus 4.8, and GPT 5.5.

Sakana also ran a beta with close to 500 early users before today's general release, which is now live with both subscription tiers and pay-as-you-go pricing for heavier workloads. The company says the value showed up less in single clever answers and more in long, grinding workflows — reading, testing, comparing evidence, revising, and eventually producing something useful. Sakana plans to keep folding new models into Fugu's pool over time, including open models and its own, which is really the whole thesis here: the system gets better as the ecosystem around it gets better, without anyone being locked into one company's roadmap.