Sakana AI, the Tokyo lab known more for research papers than products, started selling its first commercial tool on June 15. It is called Sakana Marlin, and the pitch is unusual: instead of answering in seconds, it runs for up to eight hours before handing you anything.
The company calls it a "Virtual CSO," shorthand for Chief Strategy Officer. You give Marlin one research question, sharpen the scope in a short back-and-forth, then walk away. Hours later it returns a report and a slide deck, no human input in between. The product page hosts sample outputs if you want to see what eight hours of machine reasoning actually looks like.
What you actually get
The page count is where the marketing and the fine print drift apart. Sakana's English announcement talks about reports of up to roughly 100 pages. The Japanese version is more modest, describing outputs of dozens of pages. At a press demo, reporters saw runs landing between 60 and 100 pages with 60 to 80 cited sources. So the 100-page figure is a ceiling, not an average, which is worth keeping in mind before anyone promises their boss a hundred pages by morning.
Each report comes with a main body, references, and appendices. The slides are generated with image-generation AI, which is either a nice touch or a tell, depending on how you feel about presentation decks assembled by a model that also invents diagrams.
The thing underneath
Marlin sits on top of two earlier Sakana projects. The reasoning engine is AB-MCTS, or Adaptive Branching Monte Carlo Tree Search, described in a 2025 paper and recognized with a spotlight at NeurIPS 2025. The short version: at each step the system decides whether to widen its search by generating a new candidate answer or go deeper by refining an existing one, and it can route those steps across multiple frontier models. The workflow automation comes from the AI Scientist project, which was featured in Nature.
You cannot run Marlin yourself. You can run its core algorithm, though. Sakana open-sourced AB-MCTS as TreeQuest under an Apache 2.0 license, so the search method is public even if the product is locked behind a login.
"Exceeded expectations by discovering angles we hadn't even imagined." That is a senior consultant at a Tokyo firm, quoted in Sakana's own materials, and it is exactly the kind of testimonial a vendor selects for a launch. The more interesting risk is the one nobody in the press release dwells on.
An agent reasoning unsupervised for eight hours can build a wrong assumption into page 12 and carry it through the next 88. A chatbot that hallucinates in two paragraphs is annoying. A 100-page strategy document that quietly compounds a bad premise is a harder problem to catch, especially when it arrives with citations and slides that look authoritative.
Who is paying for this
Pricing is tiered and clearly aimed at enterprises, not curious individuals. There is a pay-per-use option at 100 credits per run, with credits priced at ¥98 each. Beyond that sit Pro at ¥150,000 a month, Team at ¥400,000, and custom Enterprise deals. Sakana has also pulled in serious backers, with reported ties to MUFG and strategic investment from Citigroup, which tells you which industry it expects to convert first.
The closed beta ran from April 2026 with roughly 300 professionals across finance, consulting, and think tanks. Marlin is available now through Sakana's website. Whether anyone trusts an unsupervised agent enough to act on its conclusions, rather than just admiring the page count, is the question the next few quarters will answer.
