Sakana AI brings model orchestration and Japanese AI into products

Sakana AI develops AI models and systems in Japan, with a current product portfolio that spans model orchestration, Japanese-language APIs, chat and long-running business research. Fugu coordinates model work, Namazu specializes in Japanese and business context, and Marlin produces research reports and presentations. These are different ways to use the company’s work, with different pricing and access rules. The useful buying question is which layer you need: a model endpoint inside your software, an interactive assistant or a completed research deliverable.

01 / ProductResearch ideas now have distinct product routes

The September Fugu release introduces Fugu Max and Fugu Ultra v2 as variations of a common orchestration approach. Max emphasizes the cost-capability balance, while Ultra targets more demanding tasks. Sakana describes coordinating a pool of models rather than depending on one model for every stage. Those are vendor descriptions of the design; this blueprint does not independently validate the comparative benchmark claims.

Namazu provides a Japanese-specialized API with built-in web search and code execution. The launch identifies Moonshot’s Kimi K2.6 as its open-model foundation and describes additional adaptation for Japanese and Japanese business contexts. Sakana is therefore a distinct product and company identity, while the upstream model relationship remains relevant to understanding the offer.

Marlin is a separate business research application. Its commercial launch describes autonomous investigation that produces a substantial report and summary slides. Sakana Chat offers a direct way to interact with models without integrating an API. The September Chat update adds Fugu Max and memory; a chat session should not be treated as a substitute for the API’s deployment or billing arrangement.

02 / AudienceChoose between integration and a research deliverable

A Japanese software team may be interested in Namazu because it needs responses that preserve business terminology, appropriate register and source meaning. The evaluation should use its real documents and user requests, including mixed Japanese and English material. General fluency alone does not establish that a model handles a particular product vocabulary or a customer’s nuanced instruction.

A strategy team has a different need: Marlin can investigate public developments and produce a report to question and edit before a meeting. Its current FAQ limits the service to business use and explicitly excludes users in EU and EEA member states. That restriction is specific to Marlin; it should not be generalized to all Sakana products. Eligible teams still need to confirm availability in their location and the fit with their research needs.

Our Moonshot AI blueprint provides context for the underlying model ecosystem named in the Namazu announcement. Our Perplexity blueprint is a useful comparison for research and answer discovery. Compare the intended work and review process, not a single headline benchmark or the mere presence of web search.

03 / WorkflowProposed workflow for a bilingual market-entry question

Consider an eligible business in a supported location assessing whether a new industrial product has a plausible market in Japan. This is a proposed evaluation, not a test run by Sequenced. Give the research owner a precise question, a date boundary and a short list of decisions the evidence must inform. Separate facts such as an announced facility opening from inferences about potential demand.

Start with a narrow Marlin brief based on public information. Ask for alternative explanations, contradictory evidence and original sources. Specify which claims would change the decision and which topics should be excluded. A broad request for a complete market strategy risks producing a large document whose scope is difficult to evaluate.

Use the resulting report as an evidence packet. The Marlin update introduces Interactive Reading, which lets the reader discuss the report and move from a citation toward its supporting passage. In the proposed workflow, select the most consequential claims first. Ask what the source explicitly establishes and what the report inferred, then reopen the source when the distinction matters.

The same update describes editable PowerPoint output, including source URLs on slides. Revise the deck around the actual decision rather than presenting every research branch. A concise slide should preserve uncertainty: an announced investment can be evidence of activity without being a confirmed sales opportunity. Keep the fuller report available for colleagues who need the supporting detail.

If this task becomes part of a software product, evaluate Namazu separately through the API. Use a bilingual test set with expected terminology and explicit missing-information cases. Built-in tools can help gather information, but the application still needs controls for sources, output storage and user review. A successful Marlin report does not prove that a separately assembled Namazu workflow will behave identically.

For Fugu, compare accepted results and total usage on the same representative tasks. A coordinated system may perform several internal stages before producing a visible answer. The useful measure is the cost and review effort of an accepted result, including unsuccessful investigations, rather than the price of the final paragraph alone.

04 / PricingDo not mix token pricing with research credits

The API pricing page separates Fugu consumption, API subscription allowances and Namazu usage. It explicitly says API subscriptions do not change Sakana Chat limits. The table below selects the main distinctions; detailed tool and caching charges remain part of the current tariff.

For Fugu Ultra, the pricing documentation warns that orchestration tokens in usage-detail fields are real additional usage and enter the final bill. For Namazu, a tool loop can repeatedly pass accumulated context into the model. This means prompt length and final response length alone are incomplete cost measures. Inspect the service’s returned usage record and include tools in the estimate.

Marlin’s product FAQ says cancellation during a run still consumes credits. The approximate eight-hour research duration is a vendor guideline, not a delivery guarantee from Sequenced. Schedule the human review after the research window; a report that arrives just before a decision meeting leaves little room to inspect the most important evidence.

05 / DistinctionsOrchestration and Japanese specialization answer different needs

Fugu’s central idea is choosing and coordinating model work. Namazu’s is adaptation to language and business context. These can be valuable for different reasons, and an evaluation should avoid collapsing them into a universal best-model claim. An English coding task may reveal little about Japanese customer communication, while a fluent translation says little about a long, tool-driven investigation.

Marlin makes a further distinction by packaging research as a substantial deliverable. Its newer reading and editing features acknowledge that generating more pages can move the bottleneck to the person who must understand them. The product becomes more useful when a reviewer can challenge the evidence behind a conclusion and reshape the material for colleagues.

The ability to try models through Chat can reduce the effort of an initial qualitative assessment. Preserve the boundary between that experience and production integration, however. A business application needs repeatable inputs, an understood model identifier, monitored usage and a way to recognize incomplete work. Those responsibilities are not established merely by finding an impressive answer in a chat interface.

06 / QuestionsResolve the source, language and accounting boundaries

For research, the key limitation is source availability. Marlin’s current FAQ says it is not suited to topics with very little public information, instant-response use cases or investigations relying entirely on internal private data. An empty evidence base cannot be repaired by a longer report. Include a deliberately poorly documented topic in the evaluation and inspect whether the system preserves that uncertainty.

For Japanese-language use, ask domain specialists to review meaning and tone independently. A phrase can be grammatically natural but commercially inappropriate, or a translated qualifier can change the confidence of a claim. Keep a small set of cases where the model must preserve an ambiguity instead of smoothing it into a definitive answer.

For integration, record model versions and the full billing record for each accepted task. Sakana’s portfolio is evolving quickly, and the September releases change which models are available. Recheck the tariff and supported interface when switching versions rather than assuming that a familiar product name means unchanged behavior, latency or cost.

07 / DecisionMatch the Sakana product to the work you own

Sakana AI offers several concrete entry points, but each should be judged on its own purpose. Begin with the route closest to the desired outcome and keep the others as separate evaluation decisions. That makes both the evidence and the economics easier to understand.