ChainThink reports that on June 25, Sakana AI, a Japanese AI startup, launched its multi-agent collaborative system, Fugu Ultra, on OpenRouter.
This system dynamically orchestrates GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8 through a single interface, outperforming Anthropic’s flagship model Claude Fable 5 and Mythos Preview in academic and programming benchmarks.
Fugu Ultra is a language model specifically designed to learn coordination mechanisms, with technology derived from the Trinity and Conductor papers published at ICLR 2026. It enables fine-grained, dynamic role allocation based on context: in general knowledge Q&A, it assigns Gemini 3.1 Pro as the aggregator; in mathematical computations, it switches to GPT-5.5 as the aggregator for error correction; and in multi-turn coding scenarios, it alternates between GPT-5.5 writing code and Claude Opus 4.8 auditing and debugging.
After launching on OpenRouter, this system now offers 1 million tokens of context and a maximum output of 128,000 tokens, with support for tool calling and multimodal inputs.
The pricing is $5 per million tokens for input, $30 per million tokens for output, and $0.50 per million tokens for cache reads;
The underlying model consumption from a single API call will be included in the total cost of orchestration and output, and users can control the reasoning depth and maximum token limit via the reasoning parameter.