Sakana AI: Impressions of Fugu Ultra

I tried Sakana AI's Fugu Ultra, released today, for practical coding work.

To write the conclusion first, like an AI:

"Fugu Ultra is clearly smarter than GPT-5.5 or Opus 4.8 alone, but currently, I basically don't recommend it to others."

On a scale of 10, it feels like this:

• Fugu Ultra: 9.3 points
• GPT-5.5: 7.6 points
• Opus 4.8: 7.5 points

By the way, Fable 5 is 12 points.

The task I gave Fugu Ultra was as follows:

"Implement separate features in two different branches on the same repository. Then, perform large-scale and destructive optimization and refactoring, and resolve conflicts without regression while maintaining the implementation intent of both sides."

This was a task that was quite difficult for GPT-5.5 or Opus 4.8 alone; they would make wrong judgments midway or fall into infinite loops of repeating the same fixes, requiring frequent human intervention.

On the other hand, Fugu Ultra accurately grasped the implementation intent, implicit specifications, post-refactoring structure, and impact on existing features of both branches, and integrated them as intended.

In tasks like this that maintain overall consistency across multiple change histories, I think it is clearly stronger than standalone models. Honestly, if I hadn't experienced Fable, I would have rated it quite highly.

But if you ask if it's Fable-class, I can clearly say it's different. That's how overwhelming Fable was.

It splits tasks into appropriate sizes from reasoning, passes them to sub-agents, aggregates the results, and compiles them into the final product. It performed this orchestration very lightly.

Fugu Ultra's number of interventions isn't much different from Fable's, but the perceived speed until completion is about 3 times slower.

Fable also consumed tokens at a tremendous speed, but the processing was fast enough to match, and I had no complaints about the output quality.

On the other hand, Fugu Ultra consumes tokens quickly, yet the processing is slow. The stress of being made to wait and the stress of the usage quota decreasing at the same time was very painful.

Sakana AI explains that Fugu Ultra is a model that prioritizes the quality of complex multi-step tasks at the expense of response speed, but after running practical development including large-scale tasks for about 2 hours, I hit the limit for a 5-hour unit.

I was using the $220/month Max plan, but I felt that the purpose of Fugu Ultra and the design of the usage quota were not aligned.

To leverage Fugu Ultra's intelligence, you want to give it large and complex tasks, but large tasks are slow and consume a massive amount of the usage quota, creating a dilemma where it stops at critical moments.

Conversely, for small tasks, it's too slow and too expensive. In other words, it has currently become a very halfway model.

At this rate, it would be less stressful for a capable human to give detailed instructions to GPT-5.5 or Opus and split the development process themselves.

If it stays as it is, I won't renew my subscription, and I basically won't recommend it to others.

If the monthly fee were about half and the usage quota increased about fivefold, I might consider using it again.

However, I am very much looking forward to seeing how far it can go when Fable or next-generation GPT-class models become available and are integrated into Fugu's orchestration.

It's not that Fugu Ultra is bad.

Fable has changed AI coding.