Sakana AI's Marlin Could Replace Corporate Strategists

Most AI tools are optimizing for speed. Marlin is optimizing for something different: depth that takes time on purpose, and outputs that look like the work product of a strategy team rather than a chatbot.

That's a meaningful distinction, and it's worth understanding why.

Key Points

• Marlin is a "Virtual CSO": Sakana AI's first commercial product is a B2B-only autonomous research agent that runs continuous reasoning loops for up to eight hours, returning 100-page strategy reports and executive slide decks.
• The architecture is genuinely different: Marlin runs on Adaptive Branching Monte Carlo Tree Search, a system that treats research as a branching decision tree, dynamically choosing which of multiple frontier models to deploy for each subtask.
• The founders wrote the Transformer paper: Sakana was co-founded by Llion Jones, co-author of Google's 2017 "Attention Is All You Need" paper, and David Ha, former head of research at Stability AI. These are not hobbyists.
• Enterprise data policy is strict: Customer inputs are never used for model training without explicit opt-in consent. This is a direct play for M&A, competitive intelligence, and strategy work that consumer-grade tools cannot touch.
• Pricing reflects the enterprise positioning: A single Marlin run costs 100 credits on the pay-as-you-go tier. The Pro plan runs $935 per month. The Team plan is $2,495 per month.

What Marlin Does

The workflow is straightforward in description and unusual in practice. You give Marlin a research topic. It asks a few clarifying questions to sharpen the scope. Then you step away. For the next several hours, the system formulates its own hypotheses, navigates the web, cross-references sources, maps causal relationships within the business environment, and builds toward a final deliverable. You come back to a 100-page strategy report with appendices, citations, and a separate deck of executive slides.

Sakana describes it as "a junior strategy consultant locked in a room with a whiteboard and an internet connection." The analogy is deliberately unglamorous. This is not a tool that impresses you with speed. It is a tool that is supposed to impress you with what it found.

The use cases Sakana highlighted in its launch materials give a sense of the intended scale: generating resolution scenarios for a theoretical blockade of the Strait of Hormuz, mapping global AI regulation across jurisdictions, analyzing the return of bond vigilantes in macroeconomic markets. These are not queries. They are engagements.

The Engine Behind the Reports

Understanding what makes Marlin different from Perplexity's Deep Research or OpenAI's Deep Research requires understanding the architecture. Most AI research tools run a pipeline: search, retrieve, summarize, cite. Faster pipelines run more steps in parallel. Perplexity's Search as Code, for instance, writes dynamic retrieval scripts that run thousands of steps simultaneously. These tools complete their work in 3 to 30 minutes and return 5 to 10 page summaries.

Marlin runs on Adaptive Branching Monte Carlo Tree Search, a framework Sakana published in June 2025 and open-sourced under the Apache 2.0 license before commercializing it in Marlin. The analogy Sakana uses is chess engines: rather than looking at the board and guessing, a chess engine plays out thousands of future states and evaluates each position before committing to a move. AB-MCTS does something similar for research.

At each point in the reasoning process, the system makes a dynamic choice between two behaviors. It can go wider, spawning new hypotheses when the current line of inquiry yields diminishing returns or contradictions. Or it can go deeper, refining and auditing a promising line of analysis. A Bayesian decision framework governs which choice gets made at each node. The result is a research process that adapts to what it finds rather than following a predetermined sequence.

The multi-model layer adds another dimension. Marlin treats frontier AI models as a collective intelligence network, routing each subtask to the model best suited for it. An orchestration model delegates initial ideation to one LLM, a reasoning-heavy model audits and corrects intermediate outputs, and a writing model handles final synthesis. Sakana has not disclosed which specific models are used.

This architecture was validated in competition before it was commercialized. In early 2026, Sakana's ALE-Agent took first place in the AtCoder Heuristic Contest, a combinatorial optimization challenge, beating over 800 top-tier human programmers by autonomously rebuilding and testing hundreds of solutions over four hours.

Who Built This and Why It Matters

Sakana AI was founded in Tokyo in 2023 by Llion Jones, one of the eight co-authors of "Attention Is All You Need," the 2017 Google paper that introduced the transformer architecture underlying virtually every major AI model in use today, and David Ha, former head of research at Stability AI and a Google Brain researcher.

Jones has been publicly critical of where the industry went with his work. At a TED AI conference in late 2025, he said he was "absolutely sick" of transformers, arguing that investor pressure and the fixation on scaling monolithic models had calcified research and blocked the next real breakthrough. Sakana's design philosophy follows directly from that critique: rather than building one massive model, build networks of smaller specialized models that collaborate like a school of fish.

The company raised a Series B that pushed its valuation past $2.6 billion. Investors include Nvidia, Google, Khosla Ventures, Mitsubishi UFJ Financial Group, Citi, and Salesforce. The investor list reflects where Sakana is positioning: enterprise infrastructure, defense, and finance, not consumer AI.

The Enterprise Data Case

For any organization considering Marlin for sensitive work, the data policy is the operative question. Sakana's terms state that neither the company nor its AI service providers will use customer data for model training without explicit opt-in consent. Even with consent, data is processed to remove personally identifiable information.

This matters because the alternative is real. Most consumer-grade AI tools operate in ambiguous territory around data handling, and enterprise teams conducting M&A due diligence, unreleased competitive analyses, or proprietary market research cannot afford ambiguity. The Sakana data policy is a direct sales argument against using ChatGPT Pro or Perplexity Max for sensitive strategy work.

Feedback from the closed beta, which ran from April 2026 with approximately 300 professionals from financial institutions, consulting firms, and think tanks, pointed to two consistent themes: the outputs surfaced angles the teams had not considered, and the sourcing relied on primary research rather than recycled secondary summaries.

What This Means for Strategy and Research Teams

The competitive intelligence question this product raises is not whether Marlin replaces individual analysts. It's whether it changes the economics of the strategy function at the organizational level. If an eight-hour autonomous run returns something close to what a consulting team delivers in two weeks, the math on external strategy engagements changes significantly.

That does not mean human judgment disappears from the process. Marlin is built to inform executive decision-making, not replace it. The final act of deciding what to do with a 100-page strategy report still requires the kind of contextual judgment, relationship knowledge, and organizational understanding that no research agent currently replicates.

What it does mean is that the research and synthesis layer — the part that currently consumes enormous amounts of senior analyst time — becomes a different kind of investment. Understanding where tools like this fit into your actual workflow is increasingly a strategic question, not just a technology one. Our AI marketing and growth strategy work at Winsome is spending more time on exactly this kind of infrastructure thinking.

If you want to think through where autonomous research agents create genuine leverage in your organization versus where they create new risks, our team is ready for that conversation.