Sakana AI, a Tokyo-based startup founded by former Google researchers, has released Fugu Ultra, a language model that doesn't try to beat frontier models on raw capability — instead, it orchestrates them. Think of it as a meta-model that knows which expert to call for each sub-task, then synthesizes the result into a single answer.
What Is Fugu Ultra?
Fugu Ultra is the flagship of Sakana's new "Fugu" product line. On the surface, it behaves like any other foundation model API. You send it a prompt, you get back a response. But under the hood, Fugu is running a multi-agent orchestration system. It decides when to answer a query directly, and when to delegate sub-tasks to a pool of frontier models — including Anthropic's Claude, Google's Gemini, OpenAI's GPT-5.5, and even recursive calls to itself.
The company describes Fugu as "a full multi-agent orchestration system as a single foundation model," built on research published at ICLR 2026 (the Trinity and Conductor papers).
How It Works
Rather than brute-forcing scale through larger monolithic models, Sakana trained Fugu to specialize in model selection and delegation. When a complex query arrives:
1. Fugu analyzes the task 2. It selects which model(s) in its agent pool are best suited 3. It delegates sub-tasks, waits for results, cross-validates 4. It synthesizes everything into one coherent response
All of this happens transparently. As a developer, you call one API endpoint and get one JSON response back.
Benchmarks vs Frontier Models
According to Sakana's benchmarks, Fugu Ultra matches or exceeds Anthropic Fable 5 and Mythos Preview, Google Gemini 3.1 Pro, and GPT-5.5 on coding, reasoning, scientific, and agentic benchmarks. Notably, the technical report notes that neither Fable 5 nor Mythos Preview is in Fugu's agent pool — since they aren't publicly accessible after export controls.
The base Fugu model is tuned for low-latency everyday use (coding, chat, code review), while Fugu Ultra prioritizes answer quality on hard multi-step problems like AI research, paper reproduction, and cybersecurity analysis.
Pricing and Availability
Fugu and Fugu Ultra are available via an OpenAI-compatible API. You can sign up through Sakana's product page. Pricing details haven't been fully published yet, but the API uses standard token-based billing.
Both models support opt-out for specific agents — useful for teams with data privacy or compliance requirements.
Why Developers Should Care
The biggest selling point is no single-vendor dependency. Recent export controls on Anthropic's models show how quickly access can change. Fugu routes around provider disruptions dynamically. If one model becomes unavailable or gets restricted, Fugu picks alternatives from its pool.
For teams already using multi-model strategies with multiple API keys and routing logic, Fugu replaces custom orchestration code with one API call. For teams tied to a single provider, it offers a hedge against price hikes, outages, or regulatory shocks.
Limitations
• Fugu's performance depends on the quality of models in its pool — if all providers degrade, so does Fugu
• The company doesn't disclose which specific model handled which sub-task, which may be an issue for auditability
• It's an early product — latency on complex orchestrated queries may be higher than a single-model call
• Sakana doesn't yet publish which models are in the pool at any given time
Who Should Try It
Developers running complex multi-step AI workflows, AI researchers doing literature review and paper reproduction, and teams that want to reduce vendor dependency without building their own routing infrastructure. If you already juggle three API keys and a custom router, Fugu is worth a look.