If you have been hearing the name sakana fugu ai everywhere this week and you want a straight answer on whether the flagship Fugu Ultra is actually worth paying for, this is the review for you.
I built it into my own Agent OS, ran it head-to-head against the models I use every day, and I am going to give you the honest version: where it wins, where it does not, and exactly how to set it up.
No hype. No benchmark theatre. Just what I saw with my own eyes.
The quick verdict
Here is my Sakana Fugu review in one breath.
Sakana Fugu is a clever piece of kit: a full multi-agent panel behind a single API that quietly delivers near-frontier output for roughly a quarter of what I pay for the equivalent on Fusion.
The outputs looked genuinely nicer than what I got from the models I normally reach for, and I had it running inside my stack in about an hour.
But it is not perfect, and I am not going to pretend it is.
The standard Fugu tier is the value play for most people.
Fugu Ultra earns its keep on hard, deep, multi-step work where you genuinely need the best possible answer and you are willing to wait a beat for it.
If you are in the EU or UK, there is a deal-breaker I will get to.
Want my exact Sakana Fugu setup, the Agent OS it plugs into, and the templates I use? It is all inside the AI Profit Boardroom. Join 3,600+ members building agent stacks here.
What Sakana Fugu actually is
Sakana is a Japanese AI lab, and “Fugu” is their multi-agent orchestration system delivered through one model API.
That last bit matters, so let me slow down on it.
You do not sign up to a dozen different models, juggle a dozen keys, and write routing logic yourself.
You point your tool at one endpoint, and behind that endpoint Sakana runs a panel.
Multiple models, closed and open, compete on your prompt.
A judge then synthesises a single answer from the best of what came back.
If you have used Fusion, the shape of this will feel familiar.
It auto-selects models and delegates the work for you, so you get one clean answer without thinking about which engine did what.
The trade-off is that it is one-shot: you send the prompt, the panel deliberates, and you wait for the result rather than watching tokens stream instantly.
The two tiers: Fugu and Fugu Ultra
There are two flavours and the difference is simple.
- Fugu is the low-latency tier. Fast, snappy, built for coding work (it slots nicely into Codex-style flows) and customer-facing replies where speed matters.
- Fugu Ultra is the flagship. It is tuned for maximum answer quality on hard, multi-step problems like AI research. It costs more and it thinks for longer.
So the mental model is: reach for Fugu when you want quick and good, and reach for fugu ultra when you want the best possible answer and can spare the latency.
My hands-on Sakana Fugu review
I do not trust a model until I have made it build things, so that is what I did.
I put Sakana Fugu through a spread of real tasks: a full website, a playable maze game, a spiral galaxy simulation, and an orbit and solar-system simulation.
Then I ran the same prompts through GLM 5.2, Opus 4.8, and Fusion so I had something honest to compare against.
The headline finding: Fugu’s outputs simply looked nicer.
The website was cleaner.
The maze game felt more finished.
The galaxy and orbit sims had a polish to them that the others did not quite match out of the box.
That is a subjective call, and I want you to treat it as one, but it was consistent enough across four very different builds that I noticed it every time.
What the benchmarks say
Sakana claims Fugu matches Fable and Mythos, and the published numbers mostly back that up.
Here is where the two tiers land against Fable 5 on the benchmarks that matter to me.
| Benchmark | Fable 5 | Fugu | Fugu Ultra |
|---|---|---|---|
| Terminal Bench | 80.4 | 80.2 | 82.1 |
| Live Code Bench | — | — | ~93.2 |
| SW Bench Pro | Beats both Fugus | Behind Fable 5 | Behind Fable 5 |
Read that honestly and the picture is: Fugu Ultra is even with or slightly ahead of Fable 5 on most things, with Live Code Bench around 93.2 being a strong showing.
The one place it clearly loses is SW Bench Pro, where Fable 5 beats both Fugu tiers.
So this is not a “destroys everything” model.
It is a “trades blows with the frontier for a fraction of the price” model, which is a far more interesting thing to own.
Fugu vs Fugu Ultra: which should you pick?
This is the question everyone actually wants answered, so let me make it dead simple.
| Fugu | Fugu Ultra | |
|---|---|---|
| Best for | Speed, coding, customer-facing | Deep multi-step work, AI research |
| Latency | Low and fast | Slower, thinks longer |
| Answer quality | High | Highest |
| Coding fit | Excellent (Codex-style flows) | Good, but overkill for quick tasks |
| Price | Lower | Higher |
If you run high-volume agent loops, customer support, or coding, default to Fugu.
It is fast, cheap, and the quality is more than enough for the vast majority of jobs.
If your work is genuinely hard, the kind of thing where a wrong answer costs you real time or money, that is when you spend up for Fugu Ultra.
I would not run my whole stack on Ultra.
I would route the gnarly multi-step problems to it and let standard Fugu handle the rest.
Pros and cons of Sakana Fugu
Let me give you the balanced view, because you deserve both sides before you spend a penny.
The pros
- Price. Roughly 25% of what Fusion costs for the same prompts. That is not a rounding error, that is a different budget.
- Flat-rate subscription. Sakana offers a flat-rate plan, which is gold if you run high-volume agent loops and hate watching a pay-per-usage meter spin.
- One API, zero juggling. No signing up to individual models, no key sprawl. One endpoint, panel does the rest.
- Output quality. In my hands-on builds, Fugu’s results simply looked nicer than the alternatives.
- Easy to integrate. I had it inside my Agent OS in about an hour.
The cons
- Not in the EU or UK at launch. GDPR means European and British teams are locked out for now. For a lot of my audience, this is the deal-breaker.
- One-shot latency. The panel deliberates and you wait. Great answers, but you do not get instant streaming.
- SW Bench Pro gap. Fable 5 beats both Fugu tiers there, so it is not the strongest pick for every coding benchmark.
- Benchmark hype caution. More on this below, because it matters more than people think.
A quick word on benchmark honesty
Here is a hill I will die on.
Never take a model’s self-scored benchmarks at face value.
We just lived through the “Le Chaton Fat” hoax, where a model’s numbers turned out to be theatre, and the lesson stuck with me.
Published benchmarks are a starting point, not a verdict.
The numbers above are genuinely useful, but you should still test any model on your actual work before you trust it.
That is exactly why I run Goldie Bench, so I can see how a model performs on the kinds of tasks I care about rather than the ones a lab cherry-picked.
Do the same. Run your real prompts. Believe your own eyes.
Want me to look at your AI stack and tell you exactly where Fugu, Fusion, or something else fits? Grab a free AI SEO strategy session here.
How to use Sakana Fugu: a simple walkthrough
Setting this up is genuinely easy. Here is the path I took.
- Head to sakana.ai and read the technical report so you know what the panel is doing under the hood. Two minutes well spent.
- Grab your API access. Sakana exposes two APIs, one for each tier, so decide whether you are starting on Fugu, Fugu Ultra, or both.
- Point your tool or agent framework at the single Sakana endpoint. Because it is one API, there is no per-model wiring to do.
- Pick your tier per task. Route fast, high-volume, customer-facing, and coding jobs to Fugu. Route the hard multi-step problems to Fugu Ultra.
- Run your real prompts through it and compare against whatever you use now. Build a website, a small game, or a sim, and judge the output yourself.
- If you run a lot of volume, move onto the flat-rate subscription so your costs stop scaling with every call.
That is it. I plugged it straight into my Agent OS the same way, and the whole thing took about an hour from first read to working in my stack.
If you want the Agent OS I dropped Fugu into, with the routing already wired up so the right tier handles the right job, that is what I build with my members every week.
Sakana Fugu pricing
Pricing is where this gets genuinely exciting.
Fusion on OpenRouter is pay-per-usage and, frankly, pricey.
Sakana Fugu comes in at roughly 25% of Fusion’s cost for the same prompts.
That alone is enough to make me re-route a chunk of my workload.
On top of that, Sakana offers a flat-rate subscription.
If you run agent loops that hammer the API all day, a flat rate is the difference between a predictable bill and a heart attack at the end of the month.
So the value story is straightforward: near-frontier output, a quarter of the price, and a subscription option that rewards heavy use.
You access both APIs and the technical report at sakana.ai.
Who Sakana Fugu is for
Let me save you some time.
Buy it if you are outside the EU and UK, you run high-volume agent or coding workloads, and you want frontier-ish quality without frontier pricing. The flat-rate plan makes it a no-brainer for agent loops.
Spend up for Fugu Ultra if your work is deep, multi-step, research-grade, and a better answer is worth the wait.
Skip it for now if you are in the EU or UK, because GDPR means you cannot use it at launch, or if you specifically live in SW Bench Pro territory where Fable 5 still has the edge.
For most people building agent stacks, though, this is a serious, well-priced option that deserves a spot in your routing.
My honest take on sakana fugu ai
So, is the sakana fugu ai system worth it?
For me, yes, with eyes open.
The standard Fugu tier is a value monster, Fugu Ultra is the one I save for the hard problems, and the whole thing slid into my Agent OS in about an hour.
Just remember the caveats: test it on your own work, watch the one-shot latency, and check whether you can even access it from where you are.
Do that, and you have got a genuinely smart, genuinely cheap addition to your stack.
FAQ
What is Sakana Fugu AI?
It is a multi-agent orchestration system from the Japanese lab Sakana, accessed through one API. A panel of models competes on your prompt and a judge synthesises a single answer. It comes in two tiers: Fugu for speed and Fugu Ultra for the hardest problems.
Is Sakana Fugu Ultra worth it?
If you run deep, multi-step work or AI research and want the best possible answer, yes. For everyday coding and customer-facing work, the cheaper standard Fugu tier is the better value. Both cost around 25% of Fusion for the same prompts.
How is Sakana Fugu different from Fusion?
Both run a panel and synthesise one answer, but Fugu is roughly a quarter of the price for the same prompts and offers a flat-rate subscription. Fusion on OpenRouter is pay-per-usage and pricier.
Can I use Sakana Fugu in the EU or UK?
Not at launch. It is unavailable in the EU and UK due to GDPR, which is the biggest catch for European and British teams right now.
How do I start using Sakana Fugu?
Go to sakana.ai, grab one of the two API keys, point your tool at the single endpoint, and pick the Fugu or Fugu Ultra tier per task. There is no signing up to individual models. I had it inside my Agent OS in about an hour.
Also on my other sites
I have covered this from a few different angles. If you want more, read my takes here:
Want to learn AI for free, with a community and 1,000+ AI agents to get you started? Join the free AI Money Lab here.
About Julian
Julian Goldie runs a seven-figure SEO and link-building agency with a team of 70+, and shares AI and SEO strategies with 400K+ subscribers on YouTube and 163K followers on X. He is the author of Link Building Mastery and runs the AI Profit Boardroom, a community of 3,600+ members building real AI agent systems. He spends his days testing AI tools like Sakana Fugu so you do not have to take a lab’s word for it.
📺 Video notes + links to the tools 👉