👔 It's not smarter; it's a manager
The confusion worth clearing up right away: Fugu isn’t a new LLM with a better brain. It’s an orchestrator — a system that coordinates others Models—packaged to look like a single model. Its “intelligence” lies in managing, not knowing more. The name helps: Sakana means “fish” in Japanese, and the logo shows several little fish forming a larger fish. The big fish is the system; the little ones are the team.
Why does this change everything? Because it defines the right question. If you expect “more IQ,” you’ll be disappointed—in raw reasoning, Fugu doesn’t outperform a frontier model on its own. If you expect “better coordination,” you start asking useful questions: how much the team costs e when it delivers something the baseline doesn’t. The rest of the track answers exactly that.
🆕 New here? Three terms
- Orchestrator: a system that splits a task among several models and combines the answers — not a single model.
- Conductor: The “manager” component that decides who does each part (more details in topic 2).
- Frontier model: one of the market's leading models (Opus 4.8, GPT-5.5, Gemini…) — the "specialists" the team calls.
✓ What Fugu IS
- ✓An orchestrator delivered as 1 API
- ✓A manager that splits up and merges
- ✓A hidden team of frontier models
✗ What it is NOT
- ✗A new “smarter” model
- ✗A single brain with more IQ
- ✗Something conceptually unprecedented
🎚️ The small conductor
At the heart of Fugu is a small "manager" model: the conductor (or router). It reads your request and makes a decision before any heavy work: should I handle this alone, or assemble a team? The conductor isn’t the specialist who writes the code or essay — it decides who to send it to. It’s the first of Track 1’s two questions ("who handles each part") turned into model weights.
🔀 The conductor’s decision
- A.Solves it on its own — simple requests don’t justify assembling a team; the conductor responds directly and quickly.
- B.Calls in the team — requests it considers difficult become decomposition + delegation to the pool. This is where the minutes and costs come from.
💡 Why “small” matters
A small conductor is cheap to run—the expensive intelligence lives in the specialists it calls. The risk: if it decides to “call the team” for something a strong model could handle alone, you pay for the whole team for nothing. That’s the knot the verdict (2.2) tightens.
🐟 The secret pool of frontier models
Behind the endpoint is a pool of frontier models — the specialists. From what can be observed, this includes Opus 4.8, GPT-5.5, Gemini, and possibly others. Sakana maintains this proprietary, closed roster: you don’t choose who joins, or see who was called for each request. It’s a hidden team.
Watch out for a subtle distinction: "Opus is in the pool" is a reasonable inference, not a fact published by Sakana. Knowing how to separate what’s confirmed from what’s inferred protects you from treating marketing as truth—a skill that applies to any AI product.
🔒 What the "pool" hides from you
- Who joined: you can’t see which models were called or how many.
- Why it was included: the selection logic is internal to the conductor.
- How much each one cost: everything arrives aggregated as Fugu tokens.
🔀 Decompose → delegate → merge (1 endpoint)
Here’s Fugu’s complete cycle, and it’s the maestro–specialists–fuser anatomy of Track 1 becoming concrete. The conductor decomposes your task into parts, delegates each part to a specialist in the pool, waits for everyone, and a model merges the responses into a single result. All this behind one your API call.
Conductor (decomposes) → pool specialists in parallel (cyan) → merger (merges). You make 1 call; several happen behind the scenes—and you pay for them all.
Decomposes
The conductor breaks the request into subtasks. Good decomposition is half the work — and the part where the orchestrator can shine or get tangled up.
Delegate
Each subtask goes to a specialist in the pool. They run (ideally in parallel), each reading and writing tokens—which you pay for.
Merge
A model recombines the answers into a single result. The second question (“how do they combine?”) is handled internally, without you seeing it.
💲 Fugu vs. Fugu Ultra + pricing
The line has two products: the Fugu e o Fugu Ultra (the one tested in the field test). Fugu’s price is US$ 5 input / US$ 30 output per 1 million tokens — and note: the orchestration tokens (what the private team reads and writes internally) are billed too. Launch: June 22, 2026.
📊 Side-by-side pricing (per 1M tokens)
| Model | Input | Output | Observation |
|---|---|---|---|
| Fugu | US$ 5 | US$ 30 | billed orchestration tokens |
| Opus 4.8 | US$ 5 | US$ 25 | Single model, no team |
Same input, but Fugu’s output is more expensive—and it still multiplies across the team. The list price already tells you that the orchestrator costs more per token e consumes more tokens.
💡 The detail that runs up the bill
“Billed orchestration tokens” is the most expensive phrase in the product. You don’t pay only for the final answer: you pay for every token each pool specialist read and wrote to get there. That’s why the real bill (Trilha 2.2) came in well above what the list price suggests at first glance.
👻 What the team never saw
To round out the mechanics with the right tension, here’s an honest spoiler for the verdict you’ll study in the next module: in the 38 code-graded tests, the hidden team has never found an answer that a strong model couldn’t find on its own. Not once. You paid — in time and money — for a team whose extra work didn’t produce extra correct answers.
⚠️ The trap
The elegance of the mechanics (decompose→delegate→synthesize) is seductive. But beautiful mechanics aren’t the same as better results. The question that matters isn’t “how does the team work?” but “did the team deliver something the solo model couldn’t?” Here, the answer was no.
🔭 What comes next
In Module 2.2, you’ll see the proof with exact numbers: 38 tasks, 4 waves, 36 ties, 4.5× slower, 5× more expensive — and the decision ladder that tells you, by clear criteria, when The orchestrator finally gets its place.
Quick recall (optional): which phrase best describes what Fugu IS?
✅ Module summary
Next module:
2.2 — The verdict: 38 tests, and when it’s worth it