Company and lab guide
Sakana AI: the lab behind Fugu and its orchestration research
Sakana AI is the research lab that built Fugu. This page separates the company from the product, explains the learned-orchestration research cited for the family, and maps where Fugu, Namazu, and the developer console fit.
1. Who Sakana AI is
Sakana AI is the organization behind the Fugu model family. This site treats "Sakana AI" as the lab and vendor: the entity that owns the product page at sakana.ai, runs the developer console at console.sakana.ai, and publishes the terms, model IDs, and pricing rules that the rest of this library rechecks.
Two first-party signals point to a Japan-oriented operation. The product FAQ describes access for "users outside Japan" separately, and the same pricing page documents sakana-namazu, a Japanese-specialized model. This site records those facts without inventing founders, founding dates, funding, or a headquarters address that no linked first-party page states.
SakanaFugu.com is an independent editorial project with no affiliation, endorsement, sponsorship, or product-access agreement with Sakana AI. See About.
2. The research route: learned model orchestration
The current Fugu product page grounds the system in two ICLR 2026 papers: TRINITY and the Conductor. It describes learning how to assemble, route, and coordinate expert agents instead of prescribing a hand-designed workflow. This is an orchestration claim, not evidence that the participating models have been merged into one set of weights.
The product separates three layers: Sakana AI is the organization, Fugu is the managed coordinator, and the participating models are the worker pool. Ultra and Max use fixed pools, but their internal route is not exposed to the caller. The technical report is cited across this library at arXiv 2606.21228 for the research framing, not as a purchasing proof.
TRINITY and the Conductor
The vendor describes TRINITY as an evolved coordinator with Thinker, Worker, and Verifier roles, and the Conductor as learning natural-language coordination with reinforcement learning. These are research descriptions, not a promise of identical internal roles on every API call.
Learned coordination
The orchestration is learned, not a workflow you maintain. That is the source of both the convenience and the opacity.
Rapid refresh
Official FAQ says new public frontier models take roughly two weeks of training and evaluation before a Fugu update.
Proprietary routing
The routing map is not exposed. If you need that visibility, the answer is self-managed orchestration, not a different Fugu setting.
3. The console.sakana.ai platform
The developer console is where the product becomes operable. It hosts the two pages this site rechecks most often, plus the account and key surface an integration actually needs.
- Models documentation (console.sakana.ai/models): model IDs, endpoints, supported fields, pool behavior, and the usage shape.
- Pricing documentation (console.sakana.ai/pricing): Ultra and Max rates, Fugu billing rules, subscription tiers, and Namazu.
- API keys and billing mode: the account surface where a key is created and where pay-as-you-go is selected, which is a precondition for Cyber.
- Console settings: Fugu agent opt-out and training-data opt-out live here, not on partner doors.
The console is also why "official" has a precise meaning in this library: a statement is first-party when it appears on sakana.ai or console.sakana.ai. A partner listing is not that. See third-party access.
4. Where Fugu fits in the portfolio
Fugu is the flagship multi-agent family, but it is not the only model Sakana AI documents. The portfolio, as of the September 28, 2026 recheck, has two families worth separating.
| Family | Members | Positioning |
|---|---|---|
| Fugu | fugu, fugu-ultra, fugu-max, fugu-cyber | Multi-agent coding, reasoning, cost-performance work, and defensive security through an OpenAI-compatible API |
| Namazu | sakana-namazu (sakana-namazu-v1.0) | A separate Japanese-specialized model with its own rates and tool fees |
Keep the two families apart in every estimate. This library focuses on Fugu and does not fold Namazu rates into Fugu cost models. The four Fugu models themselves are summarized on the product overview and compared in detail on the Fugu AI page.
5. Access and availability status
Sakana AI's current availability rules have hard edges that shape where Fugu can be named in an architecture.
- EU and EEA: Fugu is not yet available there while Sakana AI works toward GDPR and EU-specific compliance. The FAQ says service is not provided to member states.
- Cyber: pay-as-you-go billing plus a reviewed access request with a use case and verified contact information.
- Partners: the product page lists OpenRouter, Vercel, opencode, Creao, and Merge; a listing is not a copy of official policy.
- Subscriptions: the $20 / $100 / $200 monthly tiers are API-Platform-only and include Fugu, Fugu Ultra, and Fugu Max. Sakana Chat is a separate, free product, so an API subscription does not change its limits.
Regional rules are recorded in EU availability; the access-gated model is in Cyber access.
6. How to distinguish Sakana AI from Fugu
The confusion costs money and time: a team that treats "Sakana AI" and "Fugu" as interchangeable will import Namazu pricing, partner-listing assumptions, or company-level claims into a model decision that needs none of them.
| Term | Category | What you should do with it |
|---|---|---|
| Sakana AI | The lab and vendor | Source of product facts, terms, and availability rules |
| Fugu | The multi-agent model family | The product you call through the API |
fugu | A model ID | The default member of the family; fugu-ultra, fugu-max, and fugu-cyber are the others |
| Namazu | A separate model family | Keep its rates out of Fugu estimates |
| console.sakana.ai | The developer platform | Where keys, billing mode, and docs live |
| SakanaFugu.com | An independent guide | Not the vendor; sends corrections to its own inbox |
If a procurement question is really "who do I pay and what do I get," the answer splits cleanly: you pay Sakana AI through its console, and you get Fugu model access with the documented billing and availability rules. Everything else is a different question.
Frequently asked questions
Does Sakana AI publish founder or funding details on the pages this site checks?
This site's first-party sources do not state those details, so this page does not repeat them. Where a page is silent, this library says silent rather than inventing a fact.
Is Sakana AI a Japanese company?
The product FAQ describes access for "users outside Japan" and the same pricing page documents a Japanese-specialized Namazu model. Those are the first-party signals this site records; it does not add a headquarters claim beyond them.
How do I verify something about Sakana AI myself?
Read sakana.ai and console.sakana.ai, then compare against the dated source-check log. Time-sensitive numbers change; the September 28, 2026 recheck added Fugu Max and moved the Ultra alias to fugu-ultra-v2.0.
Where does Fugu end and another product begin?
Fugu is the multi-agent family of fugu, fugu-ultra, fugu-max, and fugu-cyber. Namazu is a separate family. The distinction matters for pricing and for which model ID you pin.
Sources and verification boundary
- Sakana AI product page: family description, availability, and partner listing
- Sakana AI Models documentation: IDs and endpoints
- Sakana AI Pricing documentation: rates, subscriptions, and Namazu
- Sakana Fugu technical report: research framing, not a purchasing proof
Company-level facts are limited to what first-party pages state as of the September 28, 2026 recheck. This page deliberately does not invent founders, dates, or funding.