Sakana Fugu launches as a single‑API multi‑agent system that dynamically orchestrates a pool of leading large language models. The service claims frontier‑level performance on coding, reasoning and scientific benchmarks without tying users to any single vendor. Developers can embed the API into existing workflows, letting the platform handle model selection and switching automatically.

The engine relies on two recent ICLR 2026 papers—TRINITY and the Conductor—to learn coordination patterns rather than hard‑coded pipelines. TRINITY assigns Thinker, Worker and Verifier roles across turns, while the Conductor uses reinforcement learning to craft natural‑language prompts that improve inter‑agent communication. Benchmarks show both Fugu variants beating publicly available frontier models.

Two model tiers—standard Fugu and Fugu Ultra—are offered via an OpenAI‑compatible endpoint. The standard tier balances latency and throughput for everyday code review or chatbot tasks, while Fugu Ultra taps a deeper specialist pool for high‑stakes problems such as Kaggle competitions, paper reproduction and patent analysis. Early adopters report measurable gains on hard reasoning benchmarks.

Because the service is not yet available in the EU/EEA, the company is working on GDPR compliance before expanding. This restriction highlights the growing tension between powerful AI orchestration tools and regional data regulations. Users outside the blocked regions can already experiment with the API to evaluate cost‑performance advantages over managing multiple model contracts themselves.