01What Sakana Fugu API Is and How It Works
Sakana Fugu API is a full multi-agent orchestration system from Japanese AI lab Sakana. It delivers Fable 5 level outputs without needing Fable 5 itself. Instead of using one model, it uses a panel of multiple models that work together.
You ask a question, and the multi-agent panel API competes head-on. Then it synthesizes the answers into one final response. This is similar to Fusion, which also uses multiple models and a judge.
But Sakana manages model selection, delegation, and everything automatically for you.
It mixes closed and open models together. You do not need to sign up for individual models. You just get the API and it handles the rest.
02Main Features: Two API Tiers and Benchmarks
Sakana offers two API tiers. Fugu balances strong performance with low latency. It is faster and fits naturally into tools like Codex for coding. You can also use it for customer-facing tasks. Fugu Ultra is the flagship model, tuned for maximum answer quality on hard multi-step problems.
It is designed for deep tasks like AI research.
On Terminal Bench, Fable 5 scored 80.4 and 82.1, while Fugu scored 80.2 and outperformed on other benchmarks. On SWE Bench Pro, Fable 5 destroyed Fugu, but on Live Code Bench, Fugu Ultra scored 93.2, 92.9, and 89.8. Most benchmarks were even or Fugu won.
- Fugu: low latency, good for quick answers and coding.
- Fugu Ultra: max quality, best for complex problems.
- Benchmarks: mostly outperforms Fable 5, except SWE Bench Pro.
03How to Set Up and Get Started
- Go to sakana.ai and sign up for an account.
- Get the API key from the website.
- Choose between the Fugu API or the Fugu Ultra API based on your needs.
- Integrate the API into your code or agent system.
- Start sending prompts and receive synthesized answers.
You can also build Sakana into an agent operating system. In the episode, Julian built it into his system within 60 minutes of its release. The API standard is similar to Fusion, making integration easy.
04Using Sakana Fugu: Examples from the Episode
Julian tested Sakana on three tasks. First, he built a website that looked super nice. Second, he created a maze game that turned out great.
Third, he made a simulation of a galaxy with high quality outputs.
He also ran an orbit task. The simulation lets you zoom in and out, change the timescale, and speed up or slow down. It has a simulated date and calculates days per second as you move around the solar system.
Comparing side by side with GLM 5.2, Fugu's website had nicer animations, colors, and UI. The galaxy simulation from Fugu was more interesting and beautiful. Against Opus 4.8, Fugu's output was more engaging and had better design.
05Results and Comparisons: Fugu vs. Other Models
In GoldieBench tests, Fugu stood shoulder to shoulder with Fable and Miphos. The side-by-side outputs were impressive, especially for UI design. Fusion created one of the nicest websites in benchmark tests, but Fugu looked even nicer with clean animations.
Cost is a big difference. Fusion via OpenRouter is a lot more expensive. With the same prompts, Sakana costs 25% of Fusion's cost.
Fusion is API-only with pay-per-usage, while Sakana offers a flat rate plan, which is better for high volume agent loops.
Be careful of companies scoring their own benchmarks. Earlier this week, Le Chatant Fat had a viral hoax. This is why Julian recommends testing yourself on GoldieBench.
06Tips, Limits, and Mistakes to Avoid
Always test AI models yourself. Benchmarks can be misleading. Some companies score their own benchmarks, like the Le Chatant Fat hoax.
Use your own tests to see real performance.
Fugu and Fusion are one-shot. You wait for the panels to reply, but you cannot go back and forth like with a CLI. Plan your prompts carefully for the best results.
If you use high volume agent loops, choose Sakana's flat rate plan over Fusion's pay-per-usage. Fugu Ultra is more expensive, so use it only for complex tasks. Fugu is better for quick answers.
07Who Should Use Sakana Fugu and Why It Matters
Sakana Fugu is for developers and AI users who want frontier capabilities without relying on a single model like Fable 5. It is ideal for building websites, games, and simulations with high quality UI.
The future is using multiple models together to get the best outputs. Sakana manages everything automatically, saving you time and money. It works with closed and open models, making it flexible.
Julian built Sakana into his agent operating system, which lets him plug in new models easily. If a model gets taken down, you remove it. If a new model drops, you add it within an hour.
This flexibility is valuable in the fast-changing AI landscape.
