Key Takeaways

  • Sakana AI introduced Fugu and Fugu Ultra, multi-agent models that orchestrate specialized AI systems through a single API.
  • The company says Fugu Ultra reaches performance levels comparable to Anthropic's Fable 5 and Mythos Preview on demanding technical benchmarks.
  • The multi-agent design aims to offer flexibility as export controls tighten and model access becomes more variable.

Sakana AI is diverging from the industry focus on monolithic model scaling by launching a pair of models, Fugu and Fugu Ultra, that act as conductors for a pool of specialized AI agents. The Tokyo-based startup's approach addresses development teams seeking reliable performance on science, engineering, and research problems without depending on a single frontier model.

While multi-agent architectures have been discussed for years, stitching them into a unified system remains a persistent challenge. Sakana AI aims to abstract this complexity for developers. Users send a single request to one endpoint, and behind the scenes, Fugu decides whether to solve the task independently or assemble a mix of domain-specific systems, verify their outputs, and merge the final response.

A recent report from IDC highlighted a rising demand for AI systems capable of coordinating multiple tools rather than relying on a single model. Sakana AI formalizes this concept by integrating a swappable pool of models. This architecture offers resilience as export restrictions change which frontier models enterprises can access; if a provider tightens access, the orchestration layer automatically routes requests to alternative models.

Sakana AI was co-founded by an author of the original Transformer paper—the foundational architecture for almost every major model deployed today. The company leverages this background to rethink current architectural assumptions, betting that future performance gains will increasingly stem from coordination rather than pure parameter scale.

According to the company, Fugu Ultra matches Anthropic's Fable 5 and Mythos Preview on difficult engineering, scientific, and reasoning benchmarks. The startup also claims the model outperforms Gemini 3.1 Pro, Opus 4.8, and GPT-5.5 on tasks including AutoResearch, mechanical design, and financial forecasting. Multiple analysts, including researchers at the MIT Technology Review, have previously noted that complex problem solving often benefits from orchestrating multiple models specialized in distinct reasoning strategies.

The standard Fugu model focuses on speed and low latency for daily coding or conversational use cases. Fugu Ultra is designed for maximum quality on multi-step problems, including AI research, paper reproduction, cybersecurity analysis, and patent search. Both are delivered through a unified API with subscription plans for individuals and pay-as-you-go options tailored for high-volume enterprise workloads.

While engineering teams often weigh whether multi-agent systems introduce restrictive overhead, data from the CNCF indicates that organizations increasingly accept some orchestration costs in exchange for flexibility in distributed workloads. Fugu balances this by automating the selection process, providing the agility to dynamically pull in the appropriate model for each step without requiring manual developer intervention.

Fugu also features recursive capabilities, allowing it to call itself when needed to spin up internal chains of reasoning. While recursion exists in agent research, packaging it as a default API behavior rather than an advanced configuration setting shifts the maintenance burden from developers to the model itself. This eliminates the need for teams to maintain custom orchestration code.

Companies experimenting with advanced automation or technical research often find that a single model lacks sufficient depth for every step of a workflow. Fugu's approach allows enterprises to depend on orchestration layers that tap into multiple providers rather than training massive in-house models. This architecture also enables nuanced cost management, routing specific sub-tasks to cheaper or faster models where appropriate.

While large providers continue pursuing all-in-one general intelligence, Sakana AI treats model specialization and orchestration as primary features. By wrapping multi-agent orchestration to function as a single unified system, the company has moved the concept from a research prototype to a commercial product. Available through one API with flexible pricing structures, Fugu provides a practical framework for teams building systems that rely on technical reasoning, distributed model ensembles, and complex multi-step analysis amid shifting frontier model access rules.