Is Sakana Fugu Ultra Worth the Cost and Wait?

TL;DR

Fugu Ultra produced results comparable to Claude Opus 4.8 in the presented 38-task test, but it was 4.5 times slower and five times more expensive overall. Its automated routing across frontier models may benefit complex software projects, yet the reviewer found little reason to replace separate Claude Code and Codex subscriptions for routine knowledge work.

Transcript

Introducing Sakana Fugu. Our Fugu Ultra model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls. So, obviously we had to put that to the test. You can see right here in my cloud code I am running on Fugu Ultra 1 million. You can see the task list that's carrying over is about our dashboa... Read More

Key Insights

• Fugu Ultra is a multi-agent orchestration service, not a standalone language model that independently exceeds frontier models. Its single API uses a manager model to divide work, select specialist models, and combine their contributions into a final response.
• Automated delegation is the central feature of Fugu Ultra. The conductor can route writing, coding, debugging, research, factual work, and other subtasks to different frontier models based on their specialties and the complexity of the request.
• The dashboard demonstration shows that Fugu Ultra can complete a substantial one-shot development task. It generated a working YouTube dashboard with live refreshes, channel statistics, audience signals, content distribution, performance analysis, video metrics, recommendations, comments, and strategy views.
• The 38-task comparison found little quality separation between Fugu Ultra and Claude Opus 4.8. Thirty-six tasks ended in a tie, even though Fugu could use Opus 4.8 and other frontier models within its own orchestration process.
• Fugu Ultra was 4.5 times slower and five times more expensive overall in the presented comparison. These disadvantages weakened the practical case for orchestration when the resulting quality was roughly equivalent to using Claude Opus 4.8 directly.
• Fugu differs from the described OpenRouter Fusion API in how it distributes work. Fusion sends the same prompt to three models and uses a judge to merge their answers, while Fugu decomposes the task and assigns different parts to selected models.
• Fugu Ultra worked inside the Claude Code harness and could use the reviewer's existing skills and research workflows. The integration was less direct than changing an endpoint and adding an API key, and the reported context display remained at zero across many conversational rounds.
• Fugu Ultra may be more valuable for heavy software development than for ordinary knowledge work. The reviewer saw potential when multiple teams share a codebase and benefit from integrated planning, coding, and review, but did not test extensive refactoring workloads.

Questions & Answers

Q: What is Sakana Fugu Ultra?

Sakana Fugu Ultra is presented as a multi-agent system exposed through one API, rather than a new standalone large language model. A small manager model analyzes a request, breaks it into subtasks, routes those subtasks to frontier models such as Claude, GPT, and Gemini, and combines their outputs into a unified response for the user.

Q: How does Fugu Ultra coordinate multiple AI models?

Fugu Ultra uses a conductor-style manager model to decide who should perform each part of a job. It can assign separate tasks to models with different specialties, then use another language-model step to combine their responses. The approach automates a workflow that users can otherwise construct manually with main agents, sub-agents, chained outputs, and specialized workers.

Q: How did Fugu Ultra compare with Claude Opus 4.8?

The reviewer ran Fugu Ultra and Claude Opus 4.8 across 38 tests that were created with Codex to reduce bias or cross-contamination. Thirty-six of the 38 tasks ended in a tie. The reported quality was therefore roughly similar overall, while Fugu took 4.5 times longer and cost five times more during the comparison.

Q: Why was Fugu Ultra slower and more expensive?

Fugu Ultra coordinates several models and must perform extra work to decompose tasks, route subtasks, collect results, and combine them. The reviewer identified speed and money as common costs of obtaining multiple model perspectives. In the test, that process corresponded with Fugu being 4.5 times slower and five times more expensive than Claude Opus 4.8 overall.

Q: What did Fugu Ultra build from the one-shot prompt?

Fugu Ultra built a functional YouTube analytics dashboard after receiving a one-shot slash-goal prompt and working for almost an hour. The result could refresh live data, show channel statistics, display audience and distribution information, identify rising and underperforming content, provide recommendations, analyze individual video metrics, and organize views for comments, audience behavior, and strategy.

Q: How is Fugu Ultra different from OpenRouter Fusion API?

The described OpenRouter Fusion API sends the same prompt to three models simultaneously and then uses a judge to merge their answers. Fugu Ultra instead breaks a broader job into separate parts and delegates those parts to models selected for their strengths. Both approaches combine multiple perspectives, but their assignment strategies differ substantially.

Q: Can Fugu Ultra run inside Claude Code?

Fugu Ultra successfully ran inside Claude Code during the review and could use the reviewer's existing skills, research processes, and regular workflows. However, connecting it was described as more complicated than changing an endpoint and supplying an API key. The setup used special routing to Fugu's server, and the displayed context usage stayed at zero during many rounds.

Q: Who might benefit most from using Fugu Ultra?

Fugu Ultra may be most useful for people doing heavy software development, building products, or coordinating multiple teams on the same codebase. Its single API can combine planning, coding, debugging, research, and review across different frontier models. The reviewer did not find it necessary for personal knowledge work and preferred existing Codex and Claude Code subscriptions.

Summary & Key Takeaways

• Fugu Ultra is presented as a multi-agent system rather than a new standalone large language model. A manager model breaks a request into tasks, delegates them to frontier models such as Claude, GPT, and Gemini, and then combines their contributions into one response through a single API.

• The demonstration showed Fugu Ultra creating a functional YouTube analytics dashboard from a one-shot slash-goal prompt. The dashboard refreshed live data, displayed performance and audience metrics, identified rising or underperforming content, offered recommendations, and provided AI analysis alongside data for videos, comments, audience behavior, and strategy.

• Across 38 generated assessments comparing Fugu Ultra with Claude Opus 4.8, 36 tasks ended in ties. Fugu was 4.5 times slower and five times more expensive overall. The reviewer considered it solid but unnecessary for routine knowledge work, while acknowledging possible value for complex, collaborative software development.