Top AI Product

The hard part of multi-agent AI was never the models. It was the wiring — routing, planning, verifying, stitching outputs back together. Sakana Fugu, released June 22 by Tokyo’s Sakana AI, buries all of that inside a single model you call like any other.

## What Fugu actually is

Fugu isn’t a bigger LLM. It’s a language model trained to command a swappable pool of other models — including copies of itself — planning, delegating, verifying, and synthesizing behind one OpenAI-compatible API. You send a prompt; Fugu decides which expert models handle it. Two variants ship: Fugu for low-latency everyday work, which slots into Codex for coding and review, and Fugu Ultra for hard, multi-step problems like AI research, paper reproduction, and security analysis. Sakana claims Ultra trades blows with Anthropic’s Fable 5 on the toughest engineering and reasoning benchmarks.

## The API is the point

Because it’s OpenAI-compatible, pointing your endpoint at Fugu turns a single call into a coordinated multi-model run — no orchestration framework, no glue code. Teams can opt specific agents out of the pool for privacy or compliance. The research underneath, Sakana’s ICLR 2026 papers TRINITY and Conductor, is real work, not a marketing wrapper.