Sakana Fugu AI: One API, A Whole Model Panel

Sakana Fugu lets you run a whole panel of AI models behind one API — it picks the models, runs them, and fuses a single answer for you, with a flagship tier (Fugu Ultra) that reportedly matches frontier models like Fable and Mythos.

For anyone automating with AI, that's the dream: one endpoint, a full model panel, no juggling providers. Here's how Sakana Fugu AI works.

What It Is

Sakana Fugu is a multi-agent orchestration system from Japanese lab Sakana, served through one API. ("Fugu" = pufferfish; Sakana was founded by ex-Google researchers, including a co-author of "Attention Is All You Need.")

Behind the single endpoint, a panel of models — closed and open — answer your prompt and get synthesised into one result.

Why Automators Care: It Orchestrates For You

The killer feature for automation is that Sakana manages model selection and delegation automatically.

You don't sign up to each model or wire up routing. You call one API; Sakana decides which models run, executes the panel, and fuses the answer. That's a whole orchestration layer collapsed into one endpoint.

It's one-shot — you wait for the panel, rather than going back and forth like a Claude CLI loop.

Same principle as OpenRouter Fusion — which is why both slot neatly into the Agent OS.

Fugu vs Fugu Ultra

Fugu — low latency, fast. Fits coding tools (Codex) and even customer-facing automations where speed matters. (Fusion, by contrast, can take 5–10 minutes per answer.)

Fugu Ultra — the flagship, tuned for maximum quality on hard, multi-step tasks like AI research. Pricier, but the one that reportedly matches Fable and Mythos.

🔥 Want Sakana wired into your agent stack already? Inside the AI Profit Boardroom I've built Sakana and Fusion into the Agent OS alongside Claude, OpenClaw and Hermes — daily updates, 5 coaching calls a week. → Get the Agent OS + Sakana System

Benchmarks — And Why To Test Yourself

On Sakana's published numbers, Fugu is even with or ahead of Fable 5 on most benchmarks (Terminal Bench: Fugu Ultra 82.1 vs Fable 5 80.4; Live Code Bench: Fugu Ultra 93.2) — except SW Bench Pro, where Fable 5 wins.

But these are self-reported, and labs have gamed their own benchmarks before. So verify independently.

On my own Goldie Bench leaderboard, Fugu Ultra scores 8.62/10 (3 golds) — see the full scorecard and compare the real outputs against other models.

Output Quality

Running my standard prompts (website, maze game, galaxy sim, orbit sim), Fugu produced clean, well-animated, genuinely nice outputs — edging out GLM 5.2, Opus 4.8, and even Fusion in side-by-sides.

Cost vs Fusion

Fusion (OpenRouter): pay-per-use, expensive.
Sakana Fugu: ~25% of Fusion's cost for the same prompts, with a flat-rate subscription option.

For high-volume agent loops, the flat rate makes Sakana the cheaper home.

How To Use It

Sign up at sakana.ai for an API key.
Choose Fugu (fast) or Fugu Ultra (deep quality).
Call the API, or plug it into Free Claude Code or the Agent OS — it handles the panel for you.

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

Frequently Asked Questions

What is Sakana Fugu AI?

A multi-agent system from Sakana, served via one API. A panel of models answers and gets fused into a single result, with auto model selection and a premium tier, Fugu Ultra.

Can I automate with it?

Yes — it's a single API that orchestrates the whole panel for you, so it drops cleanly into agents, apps, or the Agent OS.

How is it different from Fusion?

Lower latency, ~25% of the cost, and a flat-rate subscription option — same panel-and-synthesis idea.

Fugu or Fugu Ultra?

Fugu for fast, lower-cost automations; Fugu Ultra for hard, deep, multi-step tasks.

Are the benchmarks reliable?

They're self-reported — test independently. On Goldie Bench, Fugu Ultra scores 8.62/10.

The Verdict

Sakana Fugu AI collapses multi-model orchestration into a single API — fast, cheap, and frontier-class on early tests. For automation, Sakana Fugu AI is one of the most useful drops of 2026.