Sakana Fugu: Multi-Agent Router or Fable Competitor?

The Breakthrough

Sakana Fugu functions as a learned model router and multi-agent orchestration system that dynamically routes tasks to a pool of expert models and synthesizes the final output, rather than operating as a standalone foundation model.

What Actually Worked

• The system demonstrates competitive performance on specific reasoning-heavy benchmarks, achieving scores of 82.1 on Terminal Bench 2.1 and 95.5 on GPQA Diamond.
• The orchestration layer provides a viable alternative for users seeking to bypass export controls or single-provider dependency by coordinating multiple existing frontier models.
• The architecture allows for specialized handling of complex tasks by routing them to different agents, which can outperform single-model approaches in specific research or data-analysis contexts.

Before / After

• On SWE Bench Pro, Fable 5 achieves a score of 80.0, while Fugu Ultra reaches 73.7 and standard Fugu reaches 59.0.
• On the Humanity's Last Exam benchmark, Fable 5 scores 53.3, compared to 50.0 for Fugu Ultra and 48.5 for standard Fugu.
• On Terminal Bench 2.1, Fugu Ultra scores 82.1 and Fugu scores 80.2, surpassing the Fable 5 score of 89.8 reported in the source's comparative charts.

Context

Sakana Fugu is marketed as a high-performance alternative to frontier models like Fable 5 and Mythos Preview. The system uses a learned routing mechanism to manage agent pools, aiming to provide similar capabilities without relying on a single restricted model. However, practical testing reveals that while the benchmark numbers are technically accurate, the actual output quality for creative coding, Three.js simulation, and SVG generation tasks does not consistently reach the level of top-tier foundation models. The orchestration process also introduces hidden costs, as internal verification and agent coordination tokens contribute to the final billable usage.