Sakana Fugu AI: Fable-Level Outputs, 25% The Cost

Q: How do I access the Sakana Fugu API?

Sign up at sakana.ai and get an API key for Fugu or Fugu Ultra. One key handles the whole model panel.

Sakana Fugu is a new multi-agent AI system from Japanese lab Sakana — you call one API, and behind it a panel of models (closed and open) work together and synthesise a single answer, with a premium tier called Fugu Ultra that reportedly matches frontier models like Fable and Mythos.

The headline: Fable-level outputs at roughly a quarter of the cost of Fusion — and it just dropped.

Here's what Sakana Fugu AI is, how it compares, and how to use it.

What Is Sakana Fugu AI?

Sakana is a Japanese AI lab (founded by ex-Google researchers, including a co-author of the original "Attention Is All You Need" paper). "Fugu" means pufferfish.

Sakana Fugu is their multi-agent orchestration system, served through a single model API. Instead of one model answering, a panel of models — closed and open — produce responses that get synthesised into one answer.

It comes in two tiers: Fugu (fast, low-latency) and Fugu Ultra (the flagship, tuned for maximum quality on hard problems).

If that sounds familiar, it's the same idea as OpenRouter Fusion — a panel plus a synthesis step — but from Sakana, cheaper, and with a flat-rate option.

How It Works

You send a prompt to one API. Sakana automatically handles model selection and delegation — it picks which models in the panel answer, runs them, and fuses the result.

That's the big convenience: you don't sign up to each underlying model. One API key, and Sakana manages the panel for you.

One thing to know: it's one-shot. You wait for the panel to reply, rather than going back and forth like you would in a Claude CLI session.

Fugu vs Fugu Ultra

Fugu — balances strong performance with low latency. Answers come back fast, so it fits coding tools (like Codex) and even customer-facing use. This is the big edge over Fusion, which can be slow (5–10 minutes for an answer).

Fugu Ultra — the flagship, tuned for maximum answer quality on hard, multi-step problems. Use it for deep work like AI research. More expensive, but it's the one that reportedly stands shoulder-to-shoulder with Fable and Mythos.

🔥 Want Sakana built into an agent dashboard, ready to use? Inside the AI Profit Boardroom I've already wired Sakana (and Fusion) into the Agent Operating System — paste a key and go. Daily Agent OS updates + 5 coaching calls a week. → Get the Agent OS + Sakana System

The Benchmarks (And A Warning)

On Sakana's published numbers:

Benchmark	Fable 5	Fugu	Fugu Ultra
Terminal Bench	80.4	80.2	82.1
Live Code Bench	—	92.9	93.2
SW Bench Pro	Fable wins	lower	lower

So on most benchmarks Fugu is even with or beating Fable 5 — except SW Bench Pro, where Fable 5 still wins clearly.

The warning: these are largely self-reported benchmarks. We've seen labs game their own scores before (a viral benchmark "hoax" fooled a lot of smart people recently). So don't take the numbers at face value — test it yourself.

That's exactly why I built Goldie Bench, my independent model leaderboard. On there, Fugu Ultra scores 8.62/10 across the scored tasks (3 golds) — see its full Goldie Bench scorecard and compare it head-to-head with the others on real, playable demos.

Output Quality: Side By Side

I ran the same prompts I use on every model — a website, a maze game, a galaxy simulation, an orbit sim. Fugu's outputs looked genuinely great: clean UI, nice animations, interesting design.

In side-by-sides it edged out GLM 5.2 (which was buggier), Opus 4.8 (more boring UI), and even Fusion (marginal, but Fugu's animations and design looked nicer). Even Fable 5 wasn't this strong at UI.

Cost: The Real Story

This is where Sakana Fugu gets interesting.

Fusion (via OpenRouter) is pay-per-use, API only, and pricey.
Sakana Fugu runs the same kind of prompts for roughly 25% of Fusion's cost — and offers a flat-rate subscription plan.

If you're running high-volume agent loops, that flat rate can be a game-changer versus paying per call.

How To Use Sakana Fugu

Sign up at sakana.ai and grab your API key.
Choose your tier — Fugu for fast, Fugu Ultra for hard, deep tasks.
Call the API (it handles model selection and delegation for you), or plug it into a tool like Free Claude Code or the Agent OS.

Note: at launch, some users reported Sakana wasn't available in the EU/UK — check availability in your region.

Frequently Asked Questions

What is Sakana Fugu AI?

A multi-agent AI system from Japanese lab Sakana, served via one API. A panel of models (closed and open) answer and get synthesised into a single response — with a premium tier, Fugu Ultra, that reportedly matches frontier models.

How is Sakana Fugu different from Fusion?

Same panel-plus-synthesis idea, but Fugu is faster (lower latency), costs roughly 25% of Fusion, and offers a flat-rate subscription instead of pure pay-per-use.

What's the difference between Fugu and Fugu Ultra?

Fugu is low-latency and fast, good for coding and customer-facing tools. Fugu Ultra is the flagship, tuned for maximum quality on hard multi-step problems like AI research.

Is Sakana Fugu as good as Fable 5?

On most published benchmarks it's even or ahead, except SW Bench Pro where Fable 5 wins. Benchmarks are self-reported, so test it yourself — on Goldie Bench, Fugu Ultra scores 8.62/10.

How do I access the Sakana Fugu API?

The Verdict

Sakana Fugu AI is one of the most interesting drops of 2026: frontier-level outputs from a model panel, faster and far cheaper than Fusion, with a flat-rate option for heavy use. Just verify the benchmark hype yourself — but on early tests, Sakana Fugu AI is the real deal.