Verdict: Sakana Fugu Ultra represents the transition from monolithic "giant" models to intelligent multi-agent orchestration. By coordinating a pool of specialized experts through the TRINITY and Conductor frameworks, it achieves performance parity with Anthropic’s export-banned Fable 5 and Mythos Preview. For businesses, this is the first production-ready "resilience" layer that delivers frontier-class reasoning without single-vendor dependency.
Last verified: 2026-06-24 Core Architecture: Multi-agent orchestration (Thinker-Worker-Verifier) Primary Benefit: Matches export-controlled frontier performance using accessible models. Key Benchmark: 93.2% on LiveCodeBench (Fugu Ultra).
Why is Sakana Fugu called the "New Fable 5"?
The label "The New Fable 5" stems from Sakana Fugu Ultra's ability to hit the same performance ceiling as Anthropic’s Fable 5 and Mythos Preview, which were pulled from global access in June 2026 due to US export controls [1][5]. While Fable 5 is a monolithic model with restricted weights, Fugu is a conductor—a trained orchestrator that assembles a team of available expert models to produce the same high-tier results in coding and complex reasoning [2].
How does Sakana Fugu's architecture differ from a standard chatbot?
Most AI tools are single-brain systems: you ask, and one model answers. Sakana Fugu is a system as a model. It sits on top of a swappable pool of top-tier models and dynamically decides which ones to call, when to call them, and how to synthesize their work [3].
This "Learned Orchestration" is grounded in two ICLR 2026 research papers:
- TRINITY: An evolved coordinator (~0.6B parameters) that assigns three distinct roles: Thinker (high-level strategy), Worker (concrete execution), and Verifier (checks the solution for correctness) [4].
- The Conductor: A 7B model trained via reinforcement learning to discover natural-language coordination strategies, designing focused prompts and communication patterns for the agent pool [6].
Can Sakana Fugu really handle "messy" multi-step tasks?
Yes. Because Fugu uses a Verifier role to double-check its own logic at each step, it avoids the "lost plot" syndrome common in single models during long jobs. In technical demonstrations, Fugu has successfully built complex artifacts from a single prompt, including:
- A fully styled, functional website (planning, layout, and code generation).
- A playable maze game with internal logic and wall-collision detection.
- A physical galaxy simulator with stars moving according to astronomical motion rules [2].
How does Fugu Ultra perform against the current frontier?
On Sakana's published benchmarks, Fugu Ultra leads or ties on almost every coding and reasoning test available to the public.
| Benchmark | Fugu Ultra | GPT-5.5 | Gemini 3.1 Pro | Opus 4.8 |
|---|---|---|---|---|
| SWE-Bench Pro | 73.7 | 58.6 | 54.2 | 69.2 |
| TerminalBench 2.1 | 82.1 | 78.2 | 70.3 | 74.6 |
| LiveCodeBench | 93.2 | 85.3 | 88.5 | 87.8 |
| Humanity's Last Exam | 50.0 | 41.4 | 44.4 | 49.8 |
Data sourced from Sakana AI Technical Report (June 2026) [5].
Why is multi-agent orchestration safer for your business?
The abrupt disappearance of Fable 5 and Mythos 5 in June 2026 proved that relying on a single frontier provider is a structural liability [1]. A multi-agent orchestrator like Fugu provides strategic resilience:
- Provider Agility: Fugu can swap underlying models as they are released or restricted, keeping your workflows stable.
- AI Sovereignty: As a Japanese-based system, Fugu offers an alternative for international teams facing US export-control shocks.
- Cost Efficiency: Learned coordination reduces unnecessary API calls, making multi-model results more affordable than manual "fusion" pipelines [3].
What this means for you
If you are building an AI Agent OS or scaling business automation, you should move from "prompting a model" to "orchestrating a system."
- Use Fugu (Base) for routine coding assistance and everyday code reviews.
- Use Fugu Ultra for high-stakes research, patent investigations, or complex multi-step software engineering.
- Build Model-Agnostically: Ensure your internal workflows are not hardcoded to a single model that could vanish overnight.
FAQ
Q: Is Sakana Fugu a free tool? A: No. Sakana Fugu is a paid product available through an OpenAI-compatible API. It offers tiered pricing based on the level of orchestration depth required.
Q: Does Fugu include Fable 5 in its pool? A: No. Fable 5 and Mythos Preview are currently under export control and are not publicly available for inclusion in Fugu's orchestration pool [5]. Fugu matches their performance using a team of other accessible models.
Q: Can I use Fugu through standard AI tools? A: Yes. Because Fugu uses an OpenAI-compatible API, it can be plugged into any tool that allows you to change the base URL and API key, such as Claude Code or custom Agent OS setups.
Q: Why choose Fugu over a simple router? A: A standard router simply sends a prompt to the "best" model. Fugu is an orchestrator—it breaks the task apart and has multiple models work together on the same request, which is significantly more powerful for complex tasks.
Q: Where is Sakana AI based? A: Sakana AI is based in Tokyo, Japan, and focuses on research-driven approaches to AI orchestration and evolutionary model development.
Discussion
0 comments