01What is Fugu Ultra and how does it work?

Fugu Ultra is a new Japanese AI model from Sakana that can match the performance of Fable 5, a top model. It uses a multi-agent panel API that combines answers from several AI models together.

When you ask something, Fugu Ultra sends your prompt to multiple models like GPC 5.5, GLM 5.2, or Kimi K207 at the same time. Then a judge fuses their answers into one single response. This is like having many smart AI brains work together.

Fugu Ultra is designed for top performance, but it is slower than the standard Fugu. The standard Fugu focuses on speed and low latency. In this episode, Julian Goldie tested Fugu Ultra on various tasks and shared his results.

02Key features of Fugu Ultra

Fugu Ultra has several key features that make it stand out:

  • Multi-agent panel: It uses multiple AI models in parallel to create better answers.
  • Vendor agnostic: You can opt out of specific providers when using the API.
  • OpenAI compatible: You can drop it into existing tools easily.
  • Flat-rate subscription: You can choose a subscription plan instead of paying per usage.

The model is published on frontier-level benchmarks like SWE Bench Pro and GPQA. It is also designed to be fast in responding, but Fugu Ultra takes longer for better performance.

In the episode, Julian showed examples of visual outputs like a live universe and an inner solar system. These were fully animated and looked amazing.

03How to get started with Fugu Ultra

To start using Fugu Ultra, follow these steps:

  1. Go to Sakana.ai to get access, because Fugu is not available on OpenRouter.
  2. Choose a subscription plan that fits your needs; it has a flat-rate pricing model.
  3. Check if Sakana is available in your region, as some users in Europe could not access it.
  4. Integrate the API into your tools; it is OpenAI compatible, so you can use it with existing setups.
  5. Add it to your Agent OS if you have one, to manage builds in one place.

Julian built Fugu Ultra into his Agent OS, where he can ask a board of AI anytime. He also has Fusion built in for comparison.

If you are new, you can see examples of what Fugu can build, like games and simulations.

04Using Fugu Ultra for building projects

In the episode, Julian tested Fugu Ultra by building a Voxel Temple Run game. However, the output got truncated because it hit a 16k max token ceiling, so the game did not work.

He also created other things like a live galaxy and an inner solar system, which looked beautiful and were fully animated. These examples show the visual quality Fugu Ultra can produce.

To use Fugu Ultra effectively, you need to wait 5 to 10 minutes for each response because it is one-shot. You cannot go back and forth with the model like with Claude CLI. This makes building interactive or iterative projects harder.

Julian recommends using Fugu Ultra inside an agentic operating system where you can plug in prompts and come back later. That way you can work on other things while waiting.

05Test results and comparisons from GoldieBench

Julian ran tests on Fugu Ultra using GoldieBench. It outperformed Opus 4.8 on every task except the Voxel Run game, which did not work due to truncation.

Fugu Ultra also beat Fable 5 on many benchmarks like Terminal Bench 2.1 Reasoning Tests, GPQA Diamond, and Live Codebench. This is impressive for a small AI lab.

On the leaderboard, Fugu Ultra scored number six at the time. Julian thinks it will rise once more tasks are scored, but the failed Voxel Run game brought its score down.

Here are some key comparisons:

  • vs Fable 5: Matches performance but slower response time.
  • vs Opus 4.8: Crushed it on most creations.
  • vs Fusion: Fusion is faster and pay-as-you-go, but Fugu Ultra is cheaper with flat subscription.

06Tips and limits of Fugu Ultra

Fugu Ultra has several limits you should know:

  • Slow responses: It takes 5 to 10 minutes for each reply, sometimes up to 25 minutes.
  • Token limits: It hits a 16k max token ceiling, causing truncated outputs.
  • API limits: After four tests, Julian hit the limit and could not use it for five hours, even on a $200/month plan.
  • Limited availability: Some European users cannot access Sakana due to privacy rules.

To avoid problems, do not rely on Fugu Ultra for time-sensitive projects. If you need speed, use Fusion or Opus 4.8 instead.

Julian found that the best way to use these models is in parallel. If one breaks, you have others working in the background.

07Who should use Fugu Ultra and why it matters

Fugu Ultra is for people who want high-quality AI outputs without paying per use. It is good for building visual projects like games and simulations.

It matters because it shows that smaller labs can compete with big models like Fable 5. It also offers a flat-rate subscription, which can be cheaper for heavy users.

However, if you need fast, interactive feedback, you may prefer Opus 4.8 or Fusion. Julian suggests using multiple models together to get the best results.

Overall, Fugu Ultra is a cool idea and part of the move toward multi-agent systems. It works well in an agentic operating system where you can manage all your AI tools in one place.