Fugu Ultra orchestrates multi-model reasoning through a single OpenAI-compatible API
Sakana AI released Fugu Ultra, an orchestration layer that routes subtasks across a pool of models, matching Fable and Mythos performance on most benchmarks.
Sakana AI released Fugu Ultra, an orchestration layer that splits incoming prompts into subtasks and routes them across a pool of models through a single OpenAI-compatible endpoint. Fugu itself is a trained LLM coordinator that decides whether to answer directly or delegate parts of the request to other models in the pool—including recursive calls to itself. It collects the outputs and returns a unified response. On most benchmarks, Fugu Ultra performs on par with Sakana's earlier Fable and Mythos models.
The OpenAI-compatible endpoint means existing client libraries and workflows can switch to Fugu Ultra without code changes. Fugu can delegate subtasks to fresh instances of itself, creating a tree of coordinator nodes when a problem decomposes into parallel branches. That design choice suggests Sakana expects Fugu to handle deeply nested reasoning chains, not just flat task splits. The model is available for testing at sakana.ai/fugu; Sakana has not disclosed pricing, parameter count, or whether Fugu Ultra will eventually ship as open weights.
