Sakana released a multi-agent orchestration platform named Fugu that distributes work among multiple language models via an OpenAI-compatible API. The system either handles tasks end-to-end or decomposes them and routes subtasks to models such as Anthropic, OpenAI, Qwen, GLM, and DeepSeek.

Architecture and routing

The platform selects tools automatically to match project requirements and can route subtasks across providers to exploit strengths and reduce bottlenecks. This design also enables operational workarounds for export constraints, similar to previously observed approaches in other projects.

Use cases

Examples of tasks that the system targets include optimizing training configurations for a specific GPU, forecasting market prices, and analysing and generating device designs. The orchestration layer aims to combine complementary model capabilities into coordinated workflows.

• Finding optimal training setup for a specified GPU configuration.
• Predicting market price dynamics using ensemble model inputs.
• Analysing designs and producing schematics for hardware prototypes.

Versions and costs

Sakana offers a lightweight edition, Fugu, intended for everyday coding and chat interactions, and a heavier edition, Fugu Ultra, built for multi-step analysis, cybersecurity workflows, and scientific research. In practical tests, the internal orchestration created extensive hidden call cascades that significantly increased token consumption and cost.

In one development test for a trading terminal, Fugu Ultra consumed about $0.51, while GLM-5.2 used about $0.03, a difference of approximately 17 times. This disparity highlights how automated subtask enumeration can inflate final usage bills.

Integration and interface

The product currently does not include a native web chat interface; instead it issues API keys that users can integrate into gateways and tooling such as OpenRouter and Codex. The workflow is therefore API-centric and designed for embedding into existing developer environments.

Performance notes

On live projects, Sakana reports that the orchestration performs at levels comparable with leading models, while also warning about unpredictable token overhead from internal routing logic. Teams considering the platform should weigh orchestration benefits against potential increases in consumption.