Sakana's Fugu: orchestration as a model, and why the headline price is fiction

Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 on September 11, and the most important thing to understand about them is that they are not models. Behind one OpenAI-compatible endpoint sits an orchestration engine that builds an agent scaffold on the fly for each request and routes subtasks across a pool of open-weight and specialised models, including NVIDIA's Nemotron. The architecture comes from two ICLR 2026 papers: TRINITY, an LLM coordinator trained by evolution, and Conductor, a coordination strategy learned by reinforcement learning and expressed in natural language.

What the price sheet says and what it hides

Fugu Max lists at $2 per million input tokens and $6 per million output, with cached input at $0.25 and web calls at $0.007 each. Sakana pitches that as undercutting frontier output prices by 40-60%, and on the sheet it does.

The catch is in what counts as a token. Delegation to sub-models, verification rounds, synthesis of the sub-results and recursive calls are all billed as orchestration tokens at the standard rates. Your one query becomes N internal calls, times however many verification passes the coordinator decides on, plus a synthesis step, and you pay for all of it. The per-query cost is therefore not a function of your prompt length; it is a function of how the coordinator chose to decompose the task, which you do not control and cannot predict. Coverage of the launch puts the real cost at several times the headline for some queries.

An orchestration endpoint does not sell you tokens. It sells you a decision about how many tokens to spend, made by someone else, billed to you.

It means the $2 / $6 figure is not comparable to a model's price. Compare it to the total cost of a run, including the calls you would have made yourself.

The benchmark claim

Fugu Ultra v2.0 claims best or joint-best on 5 of 8 benchmarks, with Chartography at 48.3 versus 27.3 for Opus 5 and DeepSWE at 74.3. Two things to notice. First, the comparison pool does not include Claude Fable 5 or 5.1, nor GPT-6 Astra, which are the current top of the frontier, so best means best among the models Sakana chose to list. Second, an orchestrator with verification rounds is a test-time compute system; it should beat single-pass models on benchmarks that reward checking your work, and the fair comparison is against a single model given the same token budget. Neither of those is in the launch material.

Build versus buy orchestration

The useful thing Fugu gives everyone else is a commercial baseline. If you are deciding whether to build your own router over a heterogeneous pool, here is how I would frame it.

• Buy when the task mix is broad, you do not have evals per task type, and unpredictable per-query cost is tolerable because volume is low. Fugu makes the decomposition problem someone else's.
• Build when volume is high enough that cost variance is a budget risk, when you need to know which model saw which data, or when you already route by task. A cost-routing layer plus a fixed scaffold gives you a predictable bill and an audit trail; the frontier router pattern is the same idea with a quality gate.
• Hybrid is where most teams land: a frontier model as the orchestrator over cheaper workers, which is exactly what I described in Fable 5 orchestrating ChatGPT and Gemini. You own the scaffold; the vendor owns the models.
• Whatever you choose, run a hundred representative queries, log total tokens per query including internal calls, and plot the distribution. The tail is what you are paying for.

The honest gap

I have not run Fugu on a real workload and neither has anyone whose numbers I have seen; the cost multiplier is reported from coverage of the launch, not from a measured distribution. The benchmark claims are Sakana's, on a pool Sakana selected, without a matched compute budget. The coordinator research is peer-reviewed at ICLR, which is more than most orchestration products can say, but a paper on coordination does not tell you what your bill will be. Treat the headline price as a lower bound and the benchmarks as an invitation to run your own.