Tokyo-based Sakana AI has launched Sakana Fugu, a multi-agent orchestration system that presents itself as a single foundation model through one OpenAI-compatible API endpoint. Internally, it decides whether to answer a query directly or assemble a team of specialized models from a configurable pool.
The system is itself a language model, trained to call other LLMs, delegating tasks, managing communication between sub-agents, and synthesising results into coherent answers. It can call instances of itself recursively for deeply nested reasoning chains.
Fugu comes in two tiers. The standard version balances performance and latency for everyday coding, code review, and chatbots, and allows organisations to exclude specific providers for compliance. Fugu Ultra is optimized for maximum answer quality on hard, multi-step problems, with a fixed pool and no opt-out.
Sakana Fugu builds on two ICLR 2026 papers from the company: TRINITY, a lightweight evolved coordinator that assigns Thinker, Worker, or Verifier roles across several turns, and Conductor, which uses reinforcement learning to discover natural-language coordination strategies.
Benchmark performance
Fugu Ultra tops 10 of 11 published benchmarks, matching or beating Anthropic’s Claude Opus 4.8, Google’s Gemini 3.1 Pro, and OpenAI’s GPT-5.5 across coding, reasoning, and scientific evaluations. Key results include 73.7 per cent on SWE Bench Pro (versus 69.2 per cent for Opus 4.8), 82.1 per cent on TerminalBench 2.1, and 50 per cent on Humanity’s Last Exam.
Sakana AI frames Fugu as a hedge against single-vendor dependency, noting that newer models can be folded into the pool over time. The timing, arriving weeks after the US government ordered Anthropic to suspend Claude Fable 5 and Mythos 5 under export controls, positions Fugu as a resilience play for enterprises that cannot rely on any single frontier provider.
“Fugu allows for flexible model swapping, enabling it to dynamically circumvent the impact even if specific models are restricted,” the company said.
Use cases from beta
During a 500-user beta, Fugu Ultra demonstrated capabilities including autonomously improving a GPT training recipe across 123 experiments on one H100 GPU, solving all 300 held-out Rubik’s Cube configurations in an average of 19.72 moves each, and generating a simulated trading return of 19.43 per cent across five runs on a 50-week window.
Fugu is available as a paid API and subscription service. Fugu Ultra pricing starts at US$15 (£12) per million input tokens. It is available in Japan, the US, and APAC, with EU availability pending GDPR compliance work.
Sources: Sakana AI Launches Sakana Fugu (MarkTechPost, June 22, 2026); Sakana Fugu, Multi-Agent System as a Model (Sakana AI, June 22, 2026); Sakana AI announces ‘Sakana Fugu,’ a multi-agent system (GIGAZINE, June 22, 2026)