Sovereign‑Free Routing: How Sakana AI’s Fugu Beats Claude Fable 5 Amid Geopolitical Constraints

Technical Overview

Fugu Ultra achieved a 73.7 % score on the SWE‑bench Pro software‑engineering benchmark , outperforming GPT‑5.5 and Claude Fable 5.

Core Components

TRINITY is a coordination model of roughly 0.6 B parameters plus a routing head of under 20 K parameters . It is trained with an evolutionary algorithm . For multi‑turn tasks TRINITY partitions the external model pool into three roles:

Thinker : performs high‑level planning and task decomposition.

Worker : generates code, performs mathematical derivations, or executes logical steps.

Verifier : evaluates and corrects the Worker’s output.

Conductor is a 7 B‑parameter model trained via reinforcement learning to issue natural‑language commands that schedule other models. It adapts workflow depth to task difficulty—issuing a single request for simple tasks or constructing a multi‑stage “plan‑execute‑verify‑refine” pipeline for complex engineering problems. Conductor can also recursively invoke itself to extend test‑time compute.

The architecture (see image below) keeps the underlying agent network transparent to users. Conductor receives user intent, dispatches the Thinker to outline steps, the Worker to produce code or results, and the Verifier to enforce quality, forming a pipeline analogous to multi‑process collaboration without modifying the base large‑model weights.

Geopolitical and Resource Constraints

Founders David Ha and Llion Jones lack access to large H100 clusters and cannot compete in raw compute. Instead, Sakana AI pursued two algorithmic routes:

Evolutionary Model Merging : uses evolutionary algorithms to fuse weights from multiple open‑source models without back‑propagation, similar to genetic crossover.

Sovereign‑Free Routing : treats globally available cloud models and local open‑source models as a dynamic, plug‑in compute pool, avoiding reliance on any single proprietary API.

Fugu’s orchestration provides technical hedging against supply‑chain shocks:

Dynamic disaster‑recovery and degradation : if a closed‑source node is blocked or experiences high latency, Conductor reroutes requests to domestic open‑source clusters or unrestricted cloud nodes.

Open‑model “combo‑punch” : by looping through multiple slightly weaker open‑source models and self‑correcting, Fugu can approximate or match the inference depth of top‑tier closed models.

Implications for Developers

System‑level orchestration and test‑time compute scaling can raise application‑layer AI capability beyond the limits of single‑model parameter counts. Developers outside major compute ecosystems can:

Avoid direct competition on monolithic model size, reducing capital‑intensive hardware investment.

Build an agile, fault‑tolerant “model orchestration layer” that aggregates diverse open‑source models and heterogeneous API endpoints into a distributed cluster, leveraging algorithmic scheduling to generate synergistic performance.

As the marginal returns of ever‑larger models diminish, fine‑grained routing and orchestration at the system level will become decisive for product competitiveness.