The Security & Compliance Analyst’s New Partner
Let’s be honest: most "AI research tools" are just expensive search engines with a thesaurus.
Sakana Marlin isn’t one of those.
It’s Tokyo-based Sakana AI’s first commercial product — and it does something genuinely unusual. You give it a research topic, it thinks about it for up to eight hours straight, and then hands you back a 100-page strategy report complete with executive summary slides. No human in the loop after that initial prompt.
I know what you’re thinking. Another AI hype cycle, right? But here’s what actually makes Marlin different: it doesn’t summarize existing information. It investigates. It formulates hypotheses, goes hunting for evidence across the web, verifies what it finds, and then structures everything into strategic options a board could actually debate.
The company ran a closed beta with roughly 300 professionals from financial institutions, consulting firms, think tanks, and system integrators starting in April 2026. The feedback was telling — beta testers said Marlin dug deeper than chat-based research tools they’d used before, and one cybersecurity division at a major SIer noted the reports were grounded in primary sources rather than secondary summaries.
That’s not a PowerPoint generator. That’s an autonomous researcher that happens to work faster than any human team I’ve managed.
—
How Marlin Actually Thinks: AB-MCTS and the Multi-Model Brain
The secret sauce here is a technique called AB-MCTS — that’s Alpha-Beta Monte Carlo Tree Search, which Sakana AI published as a NeurIPS 2025 Spotlight paper. If you’ve never seen it, think of it as the AI equivalent of a detective who follows multiple leads simultaneously, abandons the dead ends quickly, and doubles down on whatever looks promising.
Here’s what that looks like in practice. You submit a topic — say, "the impact of stablecoin regulation on Japan's financial markets" or "Strait of Hormuz blockade resolution scenarios." Marlin breaks that down into sub-questions. Then it runs thousands of automated hypothesis-validation cycles, each time deciding whether to dig deeper on a particular angle or cut it entirely. The AI itself determines which arguments matter and which are noise.
But here’s where it gets interesting. Marlin doesn’t rely on a single foundation model. It orchestrates multiple models — deploying a fine-tuned LLM trained on years of strategic consulting reports for one sub-task, then switching to a specialized economic forecaster model for another, then toggling to a legal precedent analyzer when the question shifts. No single model does everything. The system asks: "Which model, right now, has the highest confidence on this sub-question?" And it routes accordingly.
This is built on Sakana AI’s broader research portfolio. The company published "The AI Scientist" in Nature — work that automates the full scientific discovery process from ideation through peer review. They also developed ALE-Agent for automated algorithm engineering. Marlin is what happens when you take all of that long-horizon reasoning research and point it at business strategy instead of laboratory science.
The result is an agent that doesn’t just retrieve information — it reasons over time. Eight hours of continuous, autonomous inference. That’s not a chatbot with a long context window. That’s something closer to what a senior strategy consultant actually does when they’re on an engagement.
—
What You Actually Get: Reports, Slides, and Strategic Options
The output isn’t a blog post. It’s structured like something you’d hand to a CFO.
Marlin produces two deliverables: a detailed report (up to 100 pages) and executive summary slides. The reports include appendices, references, charts drawn from raw data rather than stock templates, and — critically — they’re organized around strategic options, not just information summaries.
This is the part that matters most. Most AI tools give you a literature review. Marlin gives you decision frameworks. It maps the causal relationships in complex business environments and structures them into options that executives can actually discuss. The research and structuring work falls to the AI; humans focus on what they’re supposedly paid for: making decisions.
The sample use cases Sakana published are revealing. One explored stablecoin and tokenized payment regulation in Japan. Another analyzed Strait of Hormuz blockade resolution scenarios — which, given where we are geopolitically, feels less like a hypothetical and more like something a real CSO would need next Tuesday. There’s also an enterprise AI agent market map and an analysis of the global AI regulation patchwork.
One beta tester from a major consulting firm put it best: "It exceeded expectations by discovering angles we hadn’t even imagined." Another from a system integrator’s cybersecurity division noted the reports were "highly convincing, grounded strictly in primary sources." A third from a consulting division at an SIer said the depth matched "review-paper level comprehensiveness."
These aren’t casual users testing a toy. These are professionals who’ve spent years producing this kind of work manually.
—
Who This Is Actually For (And Who It Isn’t)
Sakana AI is clear about the target audience: corporate strategy and business-planning teams at financial institutions and operating companies, consulting firms, think tanks, and research houses. It’s a B2B service designed for professional business use — not consumers.
The pricing reflects that. There’s a pay-per-use tier to get started (free, with the ability to buy credits at ¥98 each), then Pro at ¥150,000/month (2,000 credits — that’s 20 research runs), Team at ¥400,000/month (6,000 credits, or 60 runs), and Enterprise with custom quoting. Each research run costs 100 credits regardless of plan.
Let’s do the math that actually matters. A human analyst working 60-hour weeks on a single strategic engagement costs roughly $12,000 when you factor in overhead, burnout risk, and attrition. Marlin runs for about $200 in compute per report. And it doesn’t call in sick.
But there are real limitations worth understanding before you sign up. Marlin works with public information. If your strategic question depends entirely on internal company data, it can’t help you yet. Niche sectors with almost no public coverage? Not its strength. Applications that need second-by-second real-time tracking? Also not built for that.
Geographic availability is another consideration. The service runs outside Japan, but EU and EEA member states are excluded — likely an AI Act compliance decision. And Sakana AI explicitly states that input data won’t be used for model training without your opt-in consent, which matters if you’re feeding it sensitive strategic questions.
The beta feedback suggests the output quality has improved significantly since the closed testing phase. Sakana says they strengthened research quality, output formatting, and long-running task stability based on what 300 professionals actually told them didn’t work.
—
What This Means for the Strategy Profession
Here’s the uncomfortable truth: if Marlin can produce boardroom-ready strategic analysis in eight hours for a fraction of the cost of a human team, then "strategy consultant" is starting to look like an entry-level title.
That’s not a threat. It’s an opportunity — if your organization is willing to actually use the tool rather than just demo it to the board.
The beta testers’ feedback reveals something important. They didn’t say Marlin replaced their judgment. They said it amplified it. The AI handles the exhaustive research and structuring; humans focus on the decisions. One consultant noted that for new research, Marlin surfaced perspectives and sources they wouldn’t have thought to look for — effectively removing their own cognitive biases from the process.
This is the pattern I expect to see repeat across knowledge work. Not AI replacing analysts, but AI making senior analysts dramatically more productive by handling the research grunt work that currently consumes 60-70% of their time. The value shifts from "finding information" to "framing the right questions and making the call."
Sakana AI positions Marlin as part of a broader product lineup that includes Sakana Chat and Sakana Fugu, with more coming. The company’s conviction — backed by that Nature-published AI Scientist research and the AB-MCTS work — is that the most capable AI comes not from a single model, but from systems that reason over time and work together.
Marlin is the first commercial proof of that idea. Whether it works for your organization depends on whether you’re willing to let an AI do the thinking and keep humans in the decision seat. The beta testers say it does exactly that.
The question isn’t whether Marlin works. It’s whether your board will let you use it.
