Sakana AI Unveils 'Fugu Ultra v2' Based on Multi-Agent Orchestration

Sakana AI has unveiled Fugu Ultra v2, adopting a learned multi-agent orchestration system unlike traditional single monolithic models. This model routes tasks through a fixed pool of open and specialized models and recursively calls itself, optimizing for complex reasoning and agentic tasks.

Performance and Architecture

• How it works: An architecture that combines multiple models rather than relying on a single one, distributing and processing tasks within a fixed model pool.
• Key use cases: Designed to deliver high performance in autonomous AI agents, coding agents, and complex multi-step reasoning tasks.
• Features: Supports configurable reasoning intensity (high, xhigh, max), function calling, structured output, image and PDF input, and built-in web search.

Performance and Specifications

• Pricing: $5.00 per input token and $30.00 per output token, with cache reads at $0.50/M and web search at $10.00/1K calls. Prompts exceeding 272K tokens are billed at higher rates.
• Context: Offers a context window of 1,000,000 tokens and can output up to 128,000 tokens.
• Accessibility: Can be integrated via OpenRouter and other platforms through an OpenAI-compatible API.

Market Response and Adoption It is gaining attention on the OpenRouter platform immediately after its release. Major agent applications such as pi (94.3M tokens) and Codex (67.4M tokens) are actively utilizing this model, demonstrating its practical utility in agentic workflows.