Sakana Fugu: Turning Multi-Agent Orchestration Into a Single Model — Japan's New Track
On June 22, Tokyo-based Sakana AI released Sakana Fugu, a model that packages multi-agent orchestration into a single LLM. Fugu Ultra benchmarks against Anthropic's Fable 5 and Mythos Preview across SWE-bench, Terminal-bench, AIME 2025, GPQA, and ARC-AGI — and notably, neither of those reference models are even in Fugu's agent pool.
💡 What You Will Learn
On June 22, Tokyo-based Sakana AI released Sakana Fugu, a model that packages multi-agent orchestration into a single LLM. Fugu Ultra benchmarks against Anthropic's Fable 5 and Mythos Preview across S
📜 Table of Contents
Not Another Big Model — It's a Model That Orchestrates Other Models
Fugu is fundamentally a multi-agent orchestration system, but its killer move is packaging the entire orchestration stack into a single language model. You call one OpenAI-compatible API; behind the scenes, Fugu decides which expert model to call, how to coordinate them, and how to synthesize a trustworthy answer. It can even recursively call itself.
The underlying technology comes from two ICLR 2026 papers: Trinity (Evolved LLM Coordinator, arXiv 2512.04695) and Conductor (Natural Language Orchestration, arXiv 2512.04388). The core idea: turn orchestration from human-written rules into a learnable model behavior.
Two tiers at launch: Fugu (low-latency, everyday use, coding-focused) and Fugu Ultra (heavy reasoning, deep tasks like AutoResearch, paper reproduction, security analysis, patent research).
Performance: Benchmarks Claim Parity with Fable 5
Fugu Ultra's reported scores (from Sakana's technical report): - SWE-bench Verified: matches Fable 5 upper bound - Terminal-bench: exceeds Mythos Preview upper bound - AIME 2025 (hard math): exceeds Mythos Preview upper bound - GPQA (graduate-level QA): matches Fable 5 - ARC-AGI: exceeds Mythos Preview upper bound
The key insight: Fugu's agent pool may run Gemini 3.1 Pro, Opus 4.8, GPT-5.5-class models (different models for subtasks), but the orchestration is owned by Fugu itself. This isn't "one giant model brute-forcing it" — it's "a team of experts plus a smart conductor."
Geopolitical Angle: Fugu as "AI Sovereignty"
Sakana explicitly states in their blog post: "Anthropic Fable 5 and Mythos Preview are subject to US export controls." Translation: some of the world's best models may become unavailable to you tomorrow.
Fugu's response: a fully swappable model pool. If Gemini pulls out of Japan or Opus restricts access to East Asia, Fugu automatically routes around the blocked supplier. Sakana brands this as "AI sovereignty" — a product-level API, not just a concept.
Japan's national stake in Sakana makes sense. Founded in 2023, backed by NTT, KDDI, Sony, Toyota, Citi, and Salesforce Ventures, with strategic partnerships with Google and Datadog. This is Japan's national backup plan.
Real-World Usage: 500 Beta Users
Beta data from 500 real users: - A software engineer: Fugu Ultra flagged 20+ bugs in code review where other tools found only 3 - A security engineer: Given a broad scope, Fugu autonomously completed the full pipeline — recon, XSS/SQLi probing, auth audit, report generation — without overstepping - An enterprise platform exec: "Personality stability" in long sessions was better than GPT-5.5
Sakana also tested Japan-specific scenarios: mechanical design, Rubik's Cube solving, Japanese handwriting recognition, single-step chess, financial time series prediction. Fugu beat Gemini 3.1 Pro, Opus 4.8, and GPT-5.5 on every one.
How to Use It
Sakana Fugu is GA now. Access via OpenAI-compatible API. Subscription (daily use) + Pay-as-you-go (enterprise heavy load). Privacy-sensitive customers can explicitly remove certain models from Fugu's pool. The commercial product Sakana Marlin (deep research assistant) is in beta with SMBC and Mitsubishi UFJ.
Personal Take
Sakana Fugu's real significance isn't matching Fable 5's scores. It's turning "orchestration" from an engineering problem into a model problem. The industry has been competing on "how big is your base model." Fugu asks: what if "how to use models" itself becomes a learnable object?
This path naturally bypasses export controls. The implications for China are more direct than they appear: while everyone chases "base model parity with GPT-5.5," a more realistic track may be dynamic orchestration of different base models by scenario. Japan got there first.
Caveats: Sakana's own benchmarks aren't third-party verified. AIME 2025 and ARC-AGI are prone to cherry-picking. And Fugu Ultra's token costs will be high — each extra hop in the orchestration chain adds cost.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
