Sakana Fugu

What it does

Sakana Fugu is a multi-agent system delivered as a single model API. It dynamically orchestrates a diverse pool of powerful models (LLMs) to tackle complex, multi-step tasks. Instead of relying on hand-designed workflows, Fugu learns to assemble, route, and coordinate expert agents for each task, achieving superior performance on coding, reasoning, and scientific benchmarks. It offers two models: Fugu (balanced performance and latency) and Fugu Ultra (optimized for maximum answer quality). Both are accessible via an OpenAI-compatible API, and users can control which agents participate in the model pool to meet data, privacy, or compliance requirements.

Who it is for

Sakana Fugu is designed for developers, data scientists, and organizations that need high-quality AI outputs for complex, quality-critical workflows. It is particularly suited for coding, code review, reasoning, scientific research, Kaggle competitions, paper reproduction, cybersecurity analysis, and literature or patent investigations. Users who want to avoid vendor lock-in and leverage collective intelligence from multiple models without managing multiple APIs will find Fugu valuable. Note: Fugu is not yet available in the EU/EEA due to GDPR compliance work.

Why it matters

Fugu addresses the limitations of relying on a single frontier model. By orchestrating multiple models, it can outperform individual frontier models on rigorous benchmarks while avoiding export control risks. It reduces API complexity by providing a single endpoint that handles model selection and switching, improving cost-performance. The system is grounded in two ICLR 2026 papers (TRINITY and Conductor) that demonstrate learned model orchestration, showing that systems can learn to coordinate agents more effectively than human-designed workflows. This approach offers frontier-level performance without single-vendor dependency.

Launch signal

Sakana Fugu is launched and available for use. The website provides pricing, FAQ, and contact information. It is also available on other platforms (not specified). The product is actively being used by early users for tasks like Kaggle competitions and paper reproduction. The website includes quantitative benchmark results comparing Fugu models with frontier models like Opus 4.8, Gemini 3.1 Pro, and GPT 5.5, showing competitive or superior performance on benchmarks such as SWE Bench Pro, TerminalBench, LiveCodeBench, and GPQA-D.

Brand and naming

The name "Sakana Fugu" combines "Sakana" (Japanese for fish) and "Fugu" (Japanese for pufferfish), evoking a sense of Japanese origin and the idea of a delicate but powerful system. The tagline "One Model to Command Them All" positions Fugu as a unified orchestrator that controls multiple models, suggesting simplicity and power. The brand emphasizes collective intelligence and dynamic orchestration, differentiating from single-model APIs. The naming is memorable and distinctive, though it may require explanation for non-Japanese speakers.