Multi-Test Winner Fugu Ultra's Multi-Agent System Beats Fable 5, Goes Live on OpenRouter

According to Dynamic Insight monitoring, the AI startup Sakana AI from Japan has launched its multi-agent collaborative system, Fugu Ultra, on OpenRouter. The system, utilizing a unified interface, dynamically orchestrates GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8, outperforming Anthropic's flagship models Claude Fable 5 and Mythos Preview in tests such as academia and programming.

Fugu Ultra itself is a language model specialized in learning coordination mechanisms, with the technology stemming from the Trinity and Conductor papers published at ICLR 2026. The system leverages Gemini 3.1 Pro as an aggregator in commonsense reasoning, switches to GPT-5.5 for aggregation error correction in mathematical computations, and alternates between GPT-5.5 for multi-round encoding and Claude Opus 4.8 for auditing and debugging, achieving fine-grained dynamic role allocation.

Following the launch on OpenRouter, Fugu Ultra offers a context window of 1 million tokens and a maximum output of 128,000 tokens, supporting tool invocation and multimodal inputs. The model incurs a fee of $5 per million tokens for inputs, $30 for outputs, and $0.5 for cache reads. Due to the cumulative underlying model consumption per single API call, the final bill will encompass the overall cost of orchestration and output. Users can also control the depth of reasoning and maximum token limits through the reasoning parameter.