Sakana Fugu: Japan's AI That Beats Fable 5 by Making Models Work as a Team

Links & Resources

• Product Page
• Release Blog
• Console (Get API Key)
• Technical Report
• TRINITY Paper
• Conductor Paper
• GitHub

Japan just launched an AI that goes toe-to-toe with Anthropic's Fable 5 and Mythos Preview - the frontier models that export controls pulled away from most of the world this month. And it does it without using either of them.

Sakana Fugu is a multi-agent orchestration system from Sakana AI (Tokyo) that works like a manager. You send one request to one API. Behind the scenes, Fugu picks the best AI model for each step, splits the work, checks it, combines everything into one clean answer - and can even call itself recursively.

Launched June 22, 2026. Generally available now.

The Benchmark Table

Full numbers across coding, reasoning, science, and agentic benchmarks:

Fugu Ultra beats or matches Opus 4.8 on every single benchmark. It beats GPT-5.5 on 8 out of 11. And it stands shoulder-to-shoulder with Fable 5 and Mythos Preview - models that aren't even in its agent pool because they're not publicly accessible.

Why This Matters Right Now

Anthropic's Fable 5 and Mythos models were pulled from most of the world this month due to export controls. If your organization relied on those models, you lost access overnight.

Fugu hits that performance level using only models you can still access. It's not a single model that could be restricted - it's an orchestration layer that routes around disruption. If one provider restricts access, Fugu dynamically switches to alternatives. That's AI sovereignty in practice.

How It Works

Sakana Fugu is itself a language model - but one trained to be a coordinator, not an answerer. When you send a request:

1. Assessment - Fugu evaluates the task. If a single model can handle it, it routes directly. If the task is complex, multi-step, or benefits from multiple perspectives, it assembles a team.

2. Orchestration - Fugu selects models from its agent pool, assigns them roles, and coordinates their work. It manages delegation, communication between agents, and verification.

3. Synthesis - Results from multiple agents are combined into one reliable answer. Contradictions are resolved, gaps are filled, and the final output is quality-checked.

4. Recursion - Fugu can call itself. If a sub-task requires its own orchestration, Fugu spawns another instance of itself to manage that sub-team.

The whole system is invisible to you. You call one endpoint, you get one answer. The complexity lives behind the API.

The Research Behind It

Fugu is built on two ICLR 2026 papers:

• TRINITY - Uses a lightweight evolved coordinator to orchestrate multiple LLMs across turns, assigning Thinker, Worker, and Verifier roles adaptively
• Conductor - Trained with reinforcement learning to discover natural-language coordination strategies, designing agent communication patterns that outperform individual models

Instead of hand-designed workflows, Fugu learns non-obvious but highly efficient collaboration patterns on its own.

Two Models

Both are available through a single OpenAI-compatible API. Switch between them by changing the model string.

Fugu lets you opt specific agents out of its pool for data/privacy/compliance requirements. Fugu Ultra uses the full pool (fixed) for maximum performance.

Real-World Results

Sakana ran Fugu against three frontier models (Gemini 3.1 Pro, Opus 4.8, GPT 5.5) on real-world tasks:

AutoResearch (AI Training Optimization) - Fugu Ultra autonomously ran 123 experiments over 14 hours on a single H100, optimizing a small GPT's training recipe. It finished with the best mean BPB (0.9774), beating all three frontier baselines.

Rubik's Cube Solver - Fugu Ultra and one frontier model wrote working solvers for 300 randomly scrambled cubes. The other two crashed on execution (0/300 solved). Fugu averaged 19.72 moves per solve - never a single move worse than the best competitor.

Blindfold Chess - Fugu played 4 blindfold games, holding the entire game in memory with no board shown. Beat all three frontier models and a 2100-Elo Stockfish engine, ending every game in checkmate.

Security Assessment - Given one scoped instruction, Fugu drove a full security audit end-to-end: recon, XSS/SQLi checks, auth review, and a clean report with evidence and retest steps.

Code Review - Where other tools flagged about 3 issues, Fugu surfaced more than 20.

How to Use It

API (OpenAI-Compatible)

from openai import OpenAI

client = OpenAI(
    base_url="https://api.sakana.ai/v1",
    api_key="YOUR_API_KEY"
)

response = client.chat.completions.create(
    model="fugu-ultra-20260615",  # or "fugu-20260615"
    messages=[{"role": "user", "content": "Your complex task here"}]
)

Works with any OpenAI-compatible client - Codex, Cursor, custom tooling. No SDK migration required.

Console

Go to console.sakana.ai to get your API key and start sending requests.

Pricing

Subscription Plans

All plans include both Fugu and Fugu Ultra. Subscribe before the end of July 2026 for a free second month.

Pay-As-You-Go (Token Plan)

Fugu: When 1 agent is active, you pay that model's standard rate. When multiple agents are active, you pay a single rate based on the top-tier model involved - fees never stack.

Fugu Ultra (fugu-ultra-20260615):

Who Built This

Sakana AI is a Tokyo-based AI lab founded by former Google Brain researchers. The name means "fish" in Japanese. Their research focuses on nature-inspired AI - evolutionary algorithms, collective intelligence, and emergent systems.

The Fugu team published two ICLR 2026 papers on learned model orchestration (TRINITY and Conductor), which form the technical foundation of this product.