Architecture choice
Sakana Fugu trades routing visibility for a learned pool
Official product copy says Fugu coordinates a pool of models through one API and learns collaboration patterns instead of using a hand-written workflow. The same FAQ says you cannot see which underlying models answered a query. That pair of facts is the real comparison with a self-managed graph.
1. What you buy when you buy Fugu
You buy a single OpenAI-compatible (and now Anthropic Messages-compatible) interface that can route work across a pool. Official pages say:
- Fugu can exclude specific providers or models from its pool.
- Ultra uses a fixed full pool and routes between one and three agents.
- Routing details are proprietary and not exposed.
- New public frontier models are expected to take roughly two weeks of training and evaluation before a Fugu update.
The useful implication is operational: less workflow code, less control over the path, and a cost object that includes orchestration tokens on Ultra.
2. What you keep when you keep a graph
A self-managed design — LangGraph, a hand-written router, or a simple fallback list — lets you log every hop, pin every model, and refuse a vendor you cannot name. You also own retries, evaluation, and the failure modes. That is more engineering. It is the correct extra work when audit, data residency, or incident response requires a named worker.
3. Decision rules that follow from official constraints
| If you must | Start with | Why |
|---|---|---|
| Ship a coding or review assistant quickly outside the EU and EEA | Fugu | Official default for latency and everyday work; opt-outs exist |
| Spend more for harder multi-step quality | Ultra, then evaluate | Fixed pool; one-to-three agent routing; published rates |
| Name every model in an audit log | Self-managed | Official FAQ says Fugu will not expose that map |
| Exclude a provider but still want a managed pool | Fugu, not Ultra | Official FAQ says Ultra’s pool is fixed |
| Serve EU or EEA production users | Neither official Fugu product | Official availability restriction |
The chooser encodes the last three rows. It will recommend self-managed orchestration when you require full routing visibility, even if the workload is otherwise a Cyber candidate.
4. Hybrid is allowed; mixed accounting is not
Some teams will keep a self-managed graph for regulated workflows and use Fugu for internal research. That is coherent if the logs and billing stay separate. It is not coherent to send Ultra traffic through a partner, then apply this site’s official Ultra formula to an invoice that does not itemize orchestration fields. Compare usage objects before you merge cost reports.
Write the decision down as a constraint, not a preference: “we must name workers” or “we can accept a proprietary route.” Preferences drift. Constraints survive the next model rename.
If the only reason you want Fugu is “we heard it beats frontier models,” stop and use the evaluation playbook on your own tasks. Official scores are vendor evidence. They do not tell you whether you can accept an uninspectable route on the workflow that actually creates legal or operational risk.
Sources
- Sakana AI product page
- Sakana AI Models documentation
- Sakana Fugu technical report for the research framing, not for a purchasing proof