Sakana Fugu

Sakana AI's multi-agent system sold as a model. You call https://api.sakana.ai/v1 with the OpenAI SDK, pass model="fugu" and a reasoning effort, and what answers is an orchestrated system routing across multiple providers underneath fugu-get-started.

The model line

fugu · fugu-ultra-v1.1 (aliased fugu-ultra) · fugu-ultra-v1.0 · fugu-cyber. Effort is high / xhigh, with max on fugu-ultra-v1.1 only (accepted elsewhere, mapped down). There is no low or medium tier — the cheap end of the effort dial doesn't exist here, which fits a system whose Codex integration raises the stream idle timeout to two hours.

fugu-cyber is a named cyber-specialist tier, and the third in this spoke: Anthropic's Mythos is a cyber/biology specialist rather than a capability ceiling (claude-opus-5), and Google ships Gemini 3.5 Flash Cyber as a domain-specialized model (alphabet-q2-2026-earnings). Three labs now sell security as its own SKU and none publishes what makes a model a cyber model. The governance backdrop is in the same corpus: Anthropic's cyber-capable line drew an export ban (claude-fable-5), so this is a regulated category, not only a marketing one.

What it is, structurally

Three things stack here that this spoke usually pages separately:

1. A router. fugu routes across all supported providers by default, and you narrow the pool per API key rather than per request. That is ai-gateway behaviour — the same one-endpoint- over-many-providers layer as omniroute — except the routing isn't the product's name, the model's is.
2. An agent system. The docs concede the multi-agent architecture in their first sentence, and the two-hour idle timeout is the tell: turns run at agent scale, not chat scale.
3. A model SKU. Priced, versioned, addressed by a string, with an effort parameter as the only knob the buyer turns.

Why that matters for the spoke's map

This wiki's map assumes a model you buy from a provider, benchmarked on capability and priced per token. Fugu breaks each assumption at once. You cannot say which model answered, so a capability benchmark measures a routing policy on the day you ran it. You cannot state a per-token price for a named model, because the composition is the product. And the provider is a reseller of other providers while presenting as a lab with its own model family — the llm-provider taxonomy's "cloud reseller" and "frontier lab" categories collapsing into one endpoint.

The honest summary is that "model" has become a product boundary rather than a technical one. Anything can sit behind an OpenAI-compatible endpoint and be sold as a model, and increasingly what sits there is an agent system. Compare claude-opus-5's fast mode — a runtime toggle on identical weights — as the same trend from the other direction: the SKU stops corresponding to a set of weights.

Unknowns

Written 2026-07-28: "Which providers are in the default pool, what the routing policy optimizes for, per-token pricing, context window, and any benchmark at all. No independent evaluation exists."

Re-checked 2026-08-30 — the benchmark half was wrong when it was written. A Sakana Fugu Technical Report (arXiv 2606.21228, Yujin Tang, Edoardo Cetin, Jinglue Xu and twelve others; v1 2026-06-19, v2 2026-06-23) predates this page by five weeks and reports Fugu on SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam and CharXiv Reasoning, claiming state of the art among publicly accessible models. This page was built from the getting-started docs and never looked for the paper. Not yet read here — see the spoke's ## Growth edges.

What actually still holds, in two different senses:

• The pool is still undisclosed. The technical report describes a family of orchestrator models that "harness and amplify the capabilities of an LLM agent team" without naming the team. So the central fact about a router — what it routes to — is withheld in the primary source too, not just in the docs.
• No third-party reproduction. Every number in circulation is Sakana's. Secondary coverage reports early hands-on testers finding a gap between the benchmark claims and ordinary use within a day of launch — which is a complaint, not a measurement, and is exactly the kind of thing this spoke should not promote to evidence. The absence that matters is a lab re-running the grid.

An open-weight version of the same routing idea arrived on 2026-10-06: interfaze-1-lite puts a Qwen-based core in front of other labs' open models (OCR, Whisper, SAM 2.1, TimesFM) and runs locally.

Related

interfaze-1-lite · fugu-get-started · sakana-ai · ai-gateway · omniroute · llm-provider · llm-api-pricing · claude-opus-5 · synthesis