Toward New Business Intelligence: Ultra Deep Research Assistant "Sakana Marlin" Beta Testing Begins

Sakana AI has spent its short life splitting attention between pure research and getting that research into the hands of businesses. The company built an "AI Scientist" that automates parts of the scientific process, developed AB-MCTS to let multiple models reason together, and has been testing autonomous agents inside bank workflows. Sakana Marlin, announced today as the company's first commercial product, is where those threads meet: a business research assistant built to run on its own for hours and hand back finished work, not just answers to a chat prompt.

Give Marlin a research topic and it goes to work without further input, running for close to eight hours before delivering a structured summary deck and a report that stretches to dozens of pages. Sakana describes the target job as the kind of deep strategic dig a chief strategy officer and a small team might otherwise spend weeks on. Early internal testing, according to the company, showed the tool going deeper on information than the research features built into existing chat services. This is a step beyond Sakana Chat, the company's earlier consumer-facing product built to showcase its post-training work — Marlin is aimed squarely at professional decision-making, not casual conversation.

The underlying idea is what Sakana calls inference scaling, and the company reaches for a 19th-century economist to explain it. William Stanley Jevons noticed that more efficient steam engines didn't cut coal use — they increased it, because efficiency made burning coal more worthwhile. Sakana argues something similar is happening with AI compute: as models get more efficient at reasoning, the response isn't to use less compute, it's to throw more of it at harder problems, letting the system think longer and deeper before answering. But the company is clear that just piling on compute doesn't work for messy, real-world business questions that don't have one correct answer. Turning hours of machine thinking into actually useful insight, rather than expensive noise, is the harder problem.

Two pieces of Sakana's own research do that work inside Marlin. AB-MCTS treats reasoning as a tree search, running hundreds or thousands of trial branches to decide which lines of inquiry are worth chasing and which should be dropped, while also picking whichever underlying language model suits a given sub-task — work that earned a spotlight slot, placing it among the top roughly 10 percent of accepted papers, at NeurIPS 2025. Layered on top is the workflow logic from AI Scientist, the system Sakana and collaborators described in a Nature paper for automating the full research cycle from idea generation through to peer review. Applied to business, that means Marlin can carry a research task through hypothesis generation, evidence gathering, resolving contradictions and final structuring into a report without a person checking in along the way.

Sakana frames the bigger problem as one of speed colliding with depth and breadth. Making a good call today means digging deep on an issue, covering a wide range of scenarios, and keeping pace with fast-moving conditions — geopolitics, regulation, capital markets, technology shifts — all at once. Human teams doing comprehensive research the traditional way can take weeks or months, by which point the world has often moved on and the analysis is stale. Sakana positions Marlin as infrastructure meant to close that gap, pulling together information and structuring options so people can spend their time on judgment rather than legwork.

Beta recruitment opens today and is free to join, with Sakana casting a wide net across financial institutions, corporate strategy and planning departments, consulting firms, think tanks and research houses — anyone doing research as a regular part of the job. The company points to sample output already produced: a 61-page report tracing how Trump-era trade and industrial policy could reshape Japan's economy sector by sector, and a 78-page analysis of AI trends inside Japan's financial industry as of March 2026, which flags a roughly 3 trillion yen scale of digital investment alongside concerns about returns not keeping pace and rising AI-driven financial crime.