Sakana Fugu AI just dropped, and it is one of the most interesting model launches I have seen all year.
The pitch is simple but bold: frontier-level intelligence without having to pay for the frontier model directly.
I built it into my own Agent OS within an hour of launch and ran my usual hands-on tests, so in this article I will show you exactly what it is, how it performed against the big names, and the one honesty warning you need before you believe a single benchmark.
What is Sakana Fugu?
Sakana is a Japanese AI lab.
“Sakana Fugu” is a full multi-agent orchestration system that you access through a single model API.
Fugu means pufferfish in Japanese (河豚 or フグ), which is a nice nod to the idea of something small on the surface hiding a lot of power underneath.
Here is the part that makes it different.
Instead of you signing up to five different model providers and wiring them together yourself, Fugu runs a multi-agent panel where multiple models, both closed and open, compete head-on on your prompt.
Then a judge model synthesises all of those competing answers into one single response.
If that sounds familiar, it is very similar to Fusion.
The whole thing auto-handles model selection and delegation, so you plug into one API and you simply do not think about which model is doing the work.
One important detail: both Fugu and Fusion are one-shot.
You send the prompt, you wait for the panel to do its thing, and you get the answer back.
It is not a back-and-forth conversation like a Claude CLI session, so you set up the prompt properly and let the panel run.
Fugu vs Fugu Ultra: the two tiers explained
There are two tiers, and the difference matters for how you use them.
Fugu is the basic tier.
It is low latency and fast, which makes it a great fit for coding tools like Codex and for customer-facing work where speed matters.
Fugu Ultra is the flagship.
It is tuned for maximum answer quality on hard, multi-step problems, the kind of work you would throw at an AI research task, and it costs more.
So the mental model is easy: reach for Fugu when you want speed and a tight cost, and reach for Fugu Ultra when the problem is genuinely hard and you want the best possible answer.
Want the Agent OS and the full Sakana system already wired up for you, plus the prompts I use to test new models? That is exactly what we build together inside the AI Profit Boardroom. Come join 3,600+ members who are plugging tools like this into their own agent workflows.
How Sakana Fugu performs on benchmarks
Sakana claims Fugu matches frontier models Fable and Mythos.
That is a big claim, so let us look at the numbers they shared.
On most benchmarks, Fugu is even with or slightly beats Fable 5, with one clear exception.
| Benchmark | Fable 5 | Fugu | Fugu Ultra |
|---|---|---|---|
| Terminal Bench | 80.4 | 80.2 | 82.1 |
| SW Bench Pro | Clearly ahead | Behind | Behind |
| Live Code Bench | — | — | ~93.2 |
On Terminal Bench it is genuinely close: Fable 5 at 80.4, Fugu at 80.2, and Fugu Ultra edging ahead at 82.1.
On SW Bench Pro, Fable 5 clearly beats both Fugu and Fugu Ultra, so that is the one area where the frontier model still has a real edge.
On Live Code Bench, Fugu Ultra posts a very high score of around 93.2.
So the headline holds up reasonably well: on most benchmarks Fugu trades blows with Fable 5, and Fugu Ultra is the one you want for the hardest coding work, except where SW Bench Pro is concerned.
The honesty warning you need before you trust any of this
Here is the part most people will skip, and it is the most important thing in this whole article.
Be very wary of companies scoring their own benchmarks.
Remember the “Le Chaton Fat” benchmark hoax?
It went viral and fooled a lot of smart people who should have known better.
When a lab publishes numbers that make its own model look amazing, that is marketing until proven otherwise.
That is not me saying Sakana is lying.
It is me saying the only benchmark that matters is the one you run yourself on the prompts you actually use.
That is exactly why I built my own Goldie Bench and why I ran the same hands-on tests on Fugu that I run on every model.
Numbers on a slide are nice; outputs on your screen are the truth.
My hands-on tests with Sakana Fugu
I ran the same four prompts I use for every model test, so this is a like-for-like comparison.
A polished website.
Fugu produced a genuinely nice UI with smooth animations, the kind of thing you could ship.
A maze game.
Clean, playable, and it did what I asked first time.
A living spiral galaxy simulation.
This one is a stress test, and Fugu gave me a zoomable galaxy that actually felt alive.
An orbit and solar-system simulation.
It came with an adjustable time scale and a simulated date, which is a nice touch that shows the model thought about the detail.
I put these side by side against GLM 5.2, Opus 4.8 and Fusion.
Honestly?
Fugu’s designs and outputs looked nicer and more interesting to me across the board.
That is a subjective call, and you should run your own prompts before you take my word for it, but the gap was visible enough that I noticed it straight away.
Pricing: where Sakana Fugu gets really interesting
This is where the launch goes from “interesting” to “I am paying attention.”
Fusion, accessed via OpenRouter, is pay-per-usage and it is pricier.
Sakana Fugu can come in at roughly 25% of Fusion’s cost for the same prompts.
Read that again: a quarter of the cost, for output that, in my tests, looked at least as good.
On top of that, Sakana offers a flat-rate subscription.
If you are running high-volume agent loops, a flat rate changes the entire economics of what you can build, because you stop watching the meter on every single call.
That combination, frontier-ish quality at a fraction of the price with a flat-rate option, is the real story here.
How to access Sakana Fugu
To get started, you sign up at sakana.ai and grab the API there.
You get two APIs: one for Fugu and one for Fugu Ultra.
There is also a technical report available if you want to go deep on how the panel and judge work.
One catch you need to know: as of launch, Sakana Fugu is not available in the EU or UK because of GDPR.
If you are reading this from those regions, keep an eye on it, because availability like this tends to expand over time.
How I built Sakana Fugu into my Agent OS
Within about an hour of launch, I had Sakana plugged into my Agent OS.
It slots in just like Fusion: you plug it in, you swap models in and out freely, and you save the outputs straight into your workspace.
That is the whole point of building an Agent OS rather than chasing each shiny new tool.
When something like Fugu drops, you do not rebuild your stack.
You add one connection and keep moving.
If you want to start for free, join my FREE AI Money Lab. You will get the community, a free AI course, and the foundations you need to start building agent workflows that actually make money.
Should you use Sakana Fugu?
Here is my honest take.
If you can access it, Fugu is absolutely worth testing, especially if you are spending real money on Fusion or frontier models right now.
The cost story alone makes it worth an afternoon of your time.
Use Fugu for fast, customer-facing and coding work, use Fugu Ultra for the hard multi-step problems, and ignore the benchmark slides until you have run your own prompts.
Build it into an Agent OS rather than treating it as a one-off, and you will be ready for the next launch too.
Want me to look at your exact setup? Book a free AI SEO strategy session and we will map out where tools like Sakana Fugu fit into your business.
Frequently asked questions
What is Sakana Fugu AI?
Sakana Fugu AI is a multi-agent orchestration system from the Japanese AI lab Sakana that you access through a single model API.
Behind that one API, a panel of closed and open models competes head-on and a judge synthesises one answer, so you do not have to sign up to individual models.
What is the difference between Fugu and Fugu Ultra?
Fugu is the basic, low-latency, fast tier built for coding tools and customer-facing work.
Fugu Ultra is the flagship tier tuned for maximum answer quality on hard multi-step problems like AI research, and it is more expensive.
How much does Sakana Fugu cost?
Sakana Fugu can run at roughly 25% of the cost of Fusion for the same prompts, and Sakana also offers a flat-rate subscription which is useful for high-volume agent loops.
Fusion via OpenRouter is pay-per-usage and pricier.
Is Sakana Fugu available in the EU or UK?
No.
As of launch Sakana Fugu is not available in the EU or UK due to GDPR, though you can still read the technical report and watch how it performs.
Should I trust Sakana Fugu’s benchmarks?
Be careful with any company scoring its own benchmarks.
The Le Chaton Fat benchmark hoax went viral and fooled people, so the safest move is to test it yourself on your own prompts, which is exactly what I do with my Goldie Bench.
Also on my other sites
About Julian
I am Julian Goldie, founder of the 7-figure SEO and link-building agency Goldie Agency, with a team of more than 70 people.
I run a YouTube channel with over 400,000 subscribers and share AI and SEO strategies with more than 163,000 followers on X.
I am the author of “Link Building Mastery” and the founder of the AI Profit Boardroom, a community of more than 3,600 members who are building real businesses with AI agents.
If you want frontier-level AI output without frontier prices, Sakana Fugu AI is exactly the kind of launch I built my Agent OS to take advantage of, and I would test it yourself the moment you can get access.
📺 Video notes + links to the tools 👉
