Sakana AI has introduced Fugu, a family of orchestrator models that reopen one of the most intriguing debates in current AI: will performance improvements come from continually training larger monolithic models or from coordinating multiple specialized models through smarter multi-agent systems?
This idea isn’t new for technical teams already working with LangGraph, CrewAI, AutoGen, MCP, code agents, validators, external tools, and RAG workflows. What’s notable is that Sakana AI has taken that intuition and presented it in a technical report with measurable results. Fugu-Ultra, its high-quality variant, scores 73.7% on SWE-Bench Pro, surpassing the 69.2% attributed to Claude Opus 4.8 in the same report table. It also achieves 82.1% on Terminal Bench 2.1, compared to GPT-5.5’s 78.2% and Opus 4.8’s 74.6%.
The core message is compelling: Fugu isn’t aiming to be “another LLM” competing solely by size. It’s a model trained to decide which agent should intervene, how to break down tasks, which outputs to verify, and when to synthesize a final response. AI is starting to look less like a single enormous brain and more like a distributed system of specialists.
An orchestrator over frontier models
Sakana AI’s report defines Fugu as a family of orchestrators that leverage and amplify the capabilities of a team of LLM agents. Users interact with Fugu as if it were a single model, but internally, it routes, delegates, and coordinates tasks among several working models. The initial pool includes models like Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.
There are two main variants. Fugu is designed for interactive, low-latency use: it selects a single worker model per input, so response times can approach those of a direct call to a frontier model. Fugu-Ultra, on the other hand, prioritizes quality and can compose workflows involving multiple agents per task, at the cost of increased latency and complexity.
| System | Approach | Main Advantage | Operational Cost |
|---|---|---|---|
| Fugu | Routing to a single worker model | Low latency and dynamic selection of best agent | Similar to a direct call, with orchestration overhead |
| Fugu-Ultra | Multi-agent flows with multiple steps | Higher quality for complex tasks | More calls, increased latency, higher cost |
| Monolithic model | One model handles entire task | Simplicity of use and deployment | Can be expensive or less optimal for specialized tasks |
| Manual multi-agent | Flows designed by developer | Fine control over process | More engineering, maintenance, points of failure |
The technical difference is significant. Fugu doesn’t just perform “voting” among models or send the same question to multiple systems. In its low-latency variant, it uses a lightweight selection module based on internal states to pick the most suitable worker. Fugu-Ultra, on the other hand, generates workflows in natural language: it breaks down the task, assigns subtasks, defines which agents can see previous responses, and decides how to synthesize the final result.
Benchmarks tell part of the story
The published results are impressive but should be interpreted with caution. Sakana AI compares Fugu and Fugu-Ultra with frontier models using benchmarks like SWE-Bench Pro, Terminal Bench 2.1, LiveCodeBench Pro, GPQA Diamond, CharXiv Reasoning, and Humanity’s Last Exam. In several tests, Fugu-Ultra outperforms the individual models it employs as workers.
On SWE-Bench Pro, Fugu-Ultra scores 73.7%, versus 69.2% for Claude Opus 4.8, 58.6% for GPT-5.5, and 54.2% for Gemini 3.1. On Terminal Bench 2.1, it reaches 82.1%, with Fugu at 80.2%. In GPQA Diamond, both variants achieve 95.5%, outperforming Opus 4.8, GPT-5.5, and Gemini 3.1 in the report’s table.
| Benchmark | Fugu-Ultra | Fugu | Claude Opus 4.8 | Gemini 3.1 | GPT-5.5 |
|---|---|---|---|---|---|
| SWE-Bench Pro | 73.7 | 59.0 | 69.2 | 54.2 | 58.6 |
| Terminal Bench 2.1 | 82.1 | 80.2 | 74.6 | 70.3 | 78.2 |
| LiveCodeBench Pro | 90.8 | 87.8 | 84.8 | 82.9 | 88.4 |
| GPQA Diamond | 95.5 | 95.5 | 92.0 | 94.3 | 93.6 |
| CharXiv Reasoning | 86.6 | 85.1 | 84.2 | 83.3 | 84.1 |
| Humanity’s Last Exam | 50.0 | 47.2 | 49.8 | 44.4 | 41.4 |
The key isn’t that a “small Japanese model” has simply beaten Claude or GPT. Fugu-Ultra achieves these results precisely because it uses powerful models as components of a larger system. The breakthrough lies in coordination: selecting the right specialist, switching models during a task, and employing cross-validation when needed.
The report provides interesting examples. In programming tasks, Fugu can use GPT-5.5 as a builder and turn to Claude Opus 4.8 at critical debugging moments. For scientific problems, it can rely more on Gemini for specialized knowledge and GPT for mathematical calculations. This domain-specific adaptability is what Sakana AI presents as a new scaling approach.
Suspicion around closed models
The success of Fugu raises an uncomfortable question: how much of the performance of large closed models truly comes from the base model, and how much from the surrounding system layer?
There’s no public proof that Claude Mythos, Fable 5, GPT-5.5, or any other closed model operates exactly like Fugu underneath. To claim it as fact would go beyond available evidence. However, it’s reasonable to suspect that modern frontier systems aren’t just calls to a plain model. Products like Claude Code, Codex, or advanced agents depend on tools, memory, command execution, context retrieval, validators, internal prompts, and feedback loops.
Fugu exposes an architecture many companies already intuit: the practical capability of an LLM isn’t just a property of its weights. It’s a property of the entire system in which it operates. The report frames this as “agentic scaffolds,” frameworks that turn autoregressive models into agents capable of planning, using tools, reviewing their work, and leveraging environmental signals.
For closed-system providers, maintaining a simple interface makes commercial sense. Customers often don’t want to know if behind the scenes there’s a single model, multiple models, routing, memory, or verifiers. They want an answer. But for developers, companies, and administrators, this opacity increasingly matters because it impacts cost, security, vendor dependency, and reproducibility.
Implications for the AI market
Fugu points to a significant trend in tech: performance no longer just depends on scaling training. It can also come from better combining existing capabilities. This has technical, economic, and geopolitical consequences.
First, modularity. Systems can incorporate new worker models as they emerge, exclude providers for privacy/compliance reasons, favor local models for sensitive data, and use premium models only when justified. Sakana AI emphasizes that orchestration enables configuring agent pools based on user, provider, privacy, or compliance restrictions.
Second, efficiency. If a simple task can be handled by a cheaper model, there’s no reason to always call the most expensive one. For complex subtasks requiring advanced debugging, the appropriate specialist can intervene only then. In an economy where token costs and latency matter, this dynamic selection offers a strong advantage.
Third, accessibility. Training a frontier model requires enormous resources. Designing a good orchestration layer is tough but potentially more accessible for companies already working with multiple models, internal tools, and proprietary data. Not everyone can create a Fugu-Ultra, but many can build architectures inspired by this logic.
Fourth, complexity. Multi-agent systems aren’t magic. They add latency, token consumption, traceability issues, error management, agent contradictions, and dependency on multiple providers. A poor orchestrator can worsen outcomes instead of improving them. Fugu’s contribution lies in training that coordination, not random agent chaining.
The battle between monolithic models and swarms of agents won’t have a single winner. Some tasks will be best served by individual models. Others will benefit from coordinated specialists. In software, science, research, cybersecurity, CAD, long-form analysis, and workflows with tools, the latter option seems increasingly attractive.
Sakana AI hasn’t proven that large closed models are obsolete. It’s demonstrated something more interesting: the “model” no longer has to be the minimum unit of competition. The new unit can be the system. In that system, routing, memory, roles, tools, and verification matter as much as the size of the LLM.
Frequently Asked Questions
What is Fugu from Sakana AI?
Fugu is a family of orchestrator models that coordinate multiple frontier language models to solve tasks. Users interact with it as if it were a single model, but internally, it can select, combine, and verify responses from different agents.
Does Fugu truly outperform Claude Opus 4.8?
In some benchmarks from the technical report, Fugu-Ultra surpasses Claude Opus 4.8, especially on SWE-Bench Pro and Terminal Bench 2.1. But it does so as a multi-agent orchestration system, not as a single isolated model.
Does Fugu run Mythos or Fable 5 underneath?
No. The report indicates that Fable 5 and Mythos Preview are not part of Fugu’s agent pool because they are not publicly accessible.
What does this mean for companies using AI in production?
It means they can achieve better results by combining specialized models, tools, validators, and routings instead of relying solely on a single premium model. The key is to design the architecture well and measure performance, cost, and latency.
Sources:
Sakana AI, Sakana Fugu Technical Report, arXiv:2606.21228v1.
