🐡 Fugu Ultra
The maestro in action. Fugu isn’t a "smarter" model: it’s an orchestrator delivered as 1 API — it breaks down your task, delegates to a secret team of frontier models, and combines the results. Here you can see how it does this and, afterward, what happened when it was tested against a strong standalone model in 38 code-graded tests.
From the outside, you hit one address. Under the hood, a conductor starts a team of frontier models (cyan), waits for them all, and merges their outputs. All of Track 2 is about what that design costs — and when it’s worth it.
Path map
Detailed content
🐡 How Fugu works
It isn’t smarter — it’s a manager. Under the hood: a small conductor, a secret pool, decompose→delegate→fuse, the cost, and what the team never saw.
Fugu isn’t a new LLM with a better brain. It’s a orchestrator packaged as a model: its “intelligence” lies in coordinating others, not in knowing more.
Clears up the central confusion. Expecting “more IQ” from Fugu leads to frustration; expecting “better management” leads to the right questions—how much does the team cost, and when is it worth it?
Orchestrator as a model · manager vs. genius · coordination ≠ knowledge.
A small “manager” model (the conductor) reads the request and decides: do I solve it myself or assemble a team? It isn’t the specialist—it’s the one who delegates.
It’s the first of the two questions (who handles each part) turned into a model weight. Understanding the conductor explains why Fugu is sometimes fast and sometimes takes minutes.
Conductor/router · alone vs. team · task-based decision.
Behind the endpoint is a pool of frontier models (Opus 4.8, GPT-5.5, Gemini and possibly others). Sakana maintains this proprietary roster; you don’t choose or see who was included.
It’s where the cost and opacity come from. "Opus is in the pool" is a reasonable inference, not a published fact — knowing the difference protects you from marketing.
Proprietary pool · closed roster · cross-provider · hidden team.
Fugu’s cycle: decomposes the task into parts, delegates each one to a specialist in the pool, and merges the responses — all hidden behind a single API call.
It’s the conductor–specialists–fusion operator anatomy from Track 1, now made concrete. Watching the cycle explains why your "1 call" turns into many paid internal calls.
Decomposition · delegation · merging · 1 endpoint, N internal calls.
There’s Fugu and Fugu Ultra (the one tested). Price: US$ 5 in / US$ 30 out per 1M tokens — and orchestration tokens are billed. Launched in Jun 22, 2026.
The output price (US$ 30) is already a warning: the team costs money. Compared with Opus 4.8 (US$ 5 in / US$ 25 out), the output is more expensive — and that cost is multiplied across the whole team.
Fugu × Fugu Ultra · US$ 5 in / US$ 30 out · orchestration tokens billed.
A spoiler on the verdict: across 38 tests, the hidden team never found an answer that a strong model couldn't find on its own. Not once. You paid for the team to get answers the solo model would already have given you.
Close the module with the right tension: the mechanics are elegant, but the question "did it deliver anything extra?" needs proof. Track 2.2 provides the proof.
Zero marginal gain · pay for orchestration · promise × proof.
📊 The verdict: 38 tests, and when it's worth it
Code-graded field test: Fugu Ultra vs Opus 4.8. 36/38 tie, 4.5× slower, 5× more expensive — and the ladder for deciding when the orchestrator earns its place.
38 tasks in 4 waves of increasing difficulty, 334 code-graded calls (hidden tests), identical prompts on both sides, 3–7 runs each, partial credit for multi-rule tasks.
The methodology is what gives the result weight. “Graded by code” removes guesswork: either it passed the test or it didn’t.
4 waves · code-graded · hidden tests · partial credit · 3–7 runs.
Score: 36 ties · 2 Opus wins (soloist) · 0 Fugu wins (team). The hidden team never found an answer the soloist didn’t.
It's the result that punctures the hype: more models didn't mean more correct answers. Orchestration didn't exceed the solo ceiling here.
36/38 tie · 2 for the soloist · 0 for the team · zero marginal gain.
Total: 357 min (Fugu) vs 80 min (Opus) — 4.5× slower. For simple code: Opus 3–6s, Fugu 2–7min for the same answer.
Latency is the invisible cost of fan-out: you wait for the whole team. For interactive work, extra minutes per task add up.
357 vs 80 min · 4.5× · summed latency · team wait time.
Total: US$ 53,60 (Fugu) vs US$ 10,66 (Opus) — 5× more expensive. Most of Fugu's cost comes from the tokens the private team reads and writes (orchestration tokens).
Shows where the money goes: you pay for the whole team, not just the final answer. It’s the hidden cost of Path 1, now with an itemized bill.
US$ 53,60 vs US$ 10,66 · 5× · orchestration tokens billed.
In the 2 tasks that weren’t ties, solo won: a stats module with 6 bugs (Fugu 80% / Opus 100%) and strict Roman numeral validation (Fugu 4/5 / Opus 5/5, accepting "IIII" once).
The team not only failed to win — where there was a difference, it lost. Coordination doesn’t make you immune to slipups; sometimes it adds failure points.
Stats: 6 bugs, 80% vs. 100% · Roman: 4/5 vs. 5/5 · team failure.
A 3-step ladder: (1) Can one strong model already solve it? → use 1 model. (2) No → want 1 endpoint + hedge, and cost/wait time don't matter? → no: 1 model is still the better buy. (3) Yes → the orchestrator earns its place.
It's the practical criterion you take home. In this evaluation, every task was handled by the top of the ladder — but the ladder applies to any future decision, including your Claude Code setup.
Decision ladder · vendor hedge · 1 endpoint · cost × wait × gain.