What is Sakana Marlin?
Sakana Marlin is a research agent for businesses from Tokyo-based Sakana AI, and the company's first commercial product. It opened to paying customers on 15 June 2026 after a closed beta that began in April with about 300 professionals from banks, operating companies, consultancies and think tanks.
You give it a topic and answer a few questions that narrow the scope. Marlin then works alone for up to roughly eight hours: it forms hypotheses, searches the web, checks contradictions and returns a referenced report of several dozen pages (Sakana's English announcement says up to a hundred) with summary slides. A September update added Interactive Reading, a chat that jumps to the relevant part of a report and highlights the cited passage, and editable PowerPoint export that fits your own template. Sakana says it builds on its AB-MCTS search method and AI Scientist research; it does not name the models.
Every run costs 100 credits. Pay as you go is ¥98 a credit, Pro is ¥150,000 a month for 2,000 credits and Team ¥400,000 a month for 6,000. It is sold to businesses only and not offered in the EU or EEA.
How much does Sakana Marlin cost?
| Plan | Price | What you get |
|---|---|---|
| Pay as you go | ¥98 per credit | 100 credits per run, no monthly fee |
| Pro | ¥150,000/month | 2,000 credits (about 20 runs); extra credits ¥90 |
| Team | ¥400,000/month | 6,000 credits (about 60 runs); extra credits ¥85 |
| Enterprise | Custom | Custom credit allocation, discounted extra credits, dedicated support |
100 credits per run. Pay as you go ¥98 per credit (no monthly fee); Pro ¥150,000/month with 2,000 credits, extra ¥90 each; Team ¥400,000/month with 6,000 credits, extra ¥85 each; Enterprise custom. Card via Stripe; free trial mentioned in Sakana's September update.
Is Sakana Marlin worth it?
A strategy question like "what happens to our Japanese stablecoin plans if the rules change" usually turns into a week of an analyst's time: dozens of tabs, a spreadsheet of sources and a deck at the end.
Marlin is built to take that week off your hands. You give it the topic, answer a few scoping questions, and leave it. For up to about eight hours it forms hypotheses, searches, throws out lines that do not hold up and returns a referenced report of several dozen pages plus slides. VentureBeat notes that the deep research modes in mainstream chat apps typically finish in 3 to 30 minutes with 5 to 10 pages, so Marlin is a different kind of tool: slower, longer and aimed at a strategy team rather than a quick answer.
The September update fixed the part beta testers complained about most, which was reading the result. Interactive Reading lets you ask the report a question and jumps to the passage, and clicking a citation opens the source with the supporting sentence highlighted. Slides now export to editable PowerPoint in your company template, each with its source URLs.
The catch: you are paying a consultant's rate for work no one outside Sakana has measured. A run costs ¥9,800 on pay as you go, ¥7,500 on Pro if you use all 20 included runs, and cancelling halfway still spends the credits. Sakana does not say which models do the work, publishes no benchmarks or security certifications, and its own FAQ says Marlin is a poor fit for niche fields with little public data, anything needing live tracking, and research that depends on your internal documents. The beta testers Sakana quotes are unnamed.
Our take: treat a Marlin report like a first draft from a capable new hire. It is worth commissioning when the alternative is a junior analyst's week, and worth checking line by line before it reaches a board pack.
Where it earns its place
A corporate strategy team at a Japanese bank
Map how a regulatory change could affect a new product line
Run one topic overnight and hand the report to an analyst to verify. On Pro, ¥150,000 a month covers about 20 runs, roughly one per working day, at ¥7,500 each.
A two-person boutique consultancy
Produce a market landscape deck for a pitch without hiring a researcher
Pay as you go at ¥98 a credit puts a single run at ¥9,800. Export the slides into the client's PowerPoint template, then check every cited URL before sending.
A think-tank analyst preparing a scenario briefing
Explore several competing explanations of a geopolitical event
Marlin compares multiple hypotheses and lays out strategic options. Interactive Reading lets you ask why a conclusion was reached and see the source passage, which speeds up the fact-check.
Before you commit
- Setup
- Sign up at marlin.sakana.ai with a card, choose pay as you go or a plan, give it a topic and answer the scoping questions. Expect the report the next morning: Sakana's guideline is about eight hours a run.
- Integrations
- Exports editable PowerPoint slides with source URLs on each slide and works with any slide template. No API or connectors to internal data are published.
- Data and privacy
- Sakana says neither it nor its outside AI providers train on your inputs unless you explicitly opt in, and opted-in data has personal information removed first. Retention and security certifications are not published, the underlying models are not named, and the service is not offered in the EU or EEA. Which AI tools train on your data?
- Cost at scale
- A Team plan at ¥400,000 a month includes 6,000 credits, about 60 runs, or roughly ¥6,667 a run; extra credits are ¥85. Sakana bills monthly by card and supports Japan's Qualified Invoice System.
- Limits
- About eight hours per run, and cancelling still consumes the credits. Sakana's FAQ says it is not suited to niche fields with little public data, real-time tracking or research that relies only on private internal data.
What's good
- Runs for up to eight hours without supervision
- Referenced reports with citations you can click through to the highlighted source
- Editable PowerPoint export in your own template
- No training on your inputs unless you opt in
- Pay-as-you-go option with no monthly fee
What's not
- Models behind it are not disclosed
- No independent tests of report quality
- Credits are spent even if you cancel a run
- Weak on niche topics, live data and internal documents
- Not offered in the EU or EEA; business customers only
Best for
Strategy, corporate planning and consulting teams, especially in Japan, that commission long desk-research reports from public sources and will check the citations.
Sakana Marlin alternatives
Sakana Marlin: common questions
- Is Sakana Marlin worth it?
- Marlin does a week of desk research overnight for about ¥9,800 a run, but you are trusting hidden models and nobody outside Sakana has checked the reports. We rate it 3.6 out of 5.
- How much does Sakana Marlin cost?
- 100 credits per run. Pay as you go ¥98 per credit (no monthly fee); Pro ¥150,000/month with 2,000 credits, extra ¥90 each; Team ¥400,000/month with 6,000 credits, extra ¥85 each; Enterprise custom. Card via Stripe; free trial mentioned in Sakana's September update.
- What are the downsides of Sakana Marlin?
- Models behind it are not disclosed; no independent tests of report quality; credits are spent even if you cancel a run; weak on niche topics, live data and internal documents; not offered in the EU or EEA; business customers only.
- Who is Sakana Marlin best for?
- Strategy, corporate planning and consulting teams, especially in Japan, that commission long desk-research reports from public sources and will check the citations.
- How much does Sakana Marlin cost for a team or in production?
- A Team plan at ¥400,000 a month includes 6,000 credits, about 60 runs, or roughly ¥6,667 a run; extra credits are ¥85. Sakana bills monthly by card and supports Japan's Qualified Invoice System.
- Is Sakana Marlin safe for company data?
- Sakana says neither it nor its outside AI providers train on your inputs unless you explicitly opt in, and opted-in data has personal information removed first. Retention and security certifications are not published, the underlying models are not named, and the service is not offered in the EU or EEA.
- What are the best Sakana Marlin alternatives?
- The closest alternatives we have reviewed are Perplexity, Elicit, Manus and ChatGPT.
Answers checked Oct 26, 2026.