Sakana AI Launches 'Ultra Deep Research' Agent: 100+ Page Reports in

Tokyo-based AI startup Sakana AI has officially launched its inaugural commercial product, Sakana Marlin. Positioned as a "Virtual CSO" (Chief Strategy Officer), Marlin is an autonomous, business-to-business research agent engineered for profound, long-horizon reasoning, a deliberate departure from the instantaneous text generation prevalent in most modern chatbots.

Setting itself apart from the current AI landscape, Marlin operates continuous, self-governing reasoning loops for up to eight hours. This extended processing time allows it to deliver deeply researched, well-cited strategy reports and executive slide decks, often exceeding 100 pages in length. The platform is now available via Sakana AI's website, with tiered pricing designed exclusively for enterprise clients, including corporations, financial institutions, and think tanks.

A New Paradigm in Enterprise AI Research

The generative AI industry has largely prioritized speed over the past two years, with the ability to create content in milliseconds becoming the norm. However, the enterprise sector is rapidly shifting its focus from superficial, rapid generation to methodical, in-depth reasoning. With Marlin, the emphasis for businesses is no longer on how quickly an AI can respond, but on the depth of its analytical capabilities.

The workflow for Sakana Marlin is distinctly different from typical large language model (LLM) interactions. Instead of intricate prompt engineering, users simply provide a core research topic. Following a brief initial discussion to refine the scope, the human user disengages, allowing Marlin to function as a self-contained digital strategy unit. Over several hours, it independently formulates hypotheses, gathers web data, cross-references sources for verification, and analyzes complex business dynamics to identify strategic insights.

Think of Marlin less as a search engine and more as a dedicated strategy consultant. You provide a strategic question in the morning, and by the end of the day, the system delivers a comprehensive, professional-grade portfolio. This output is not a generic text block but a structured set of strategic options, complete with an executive summary, slides, appendices, and detailed references. Real-world applications showcased include resolution scenarios for the Strait of Hormuz, global AI regulation mapping, and analysis of macroeconomic trends.

The Engine Behind Long-Horizon Reasoning

Marlin represents the commercial application of Sakana AI's two years of extensive laboratory advancements. Its exploration engine is powered by the company's prior research breakthrough, Adaptive Branching Monte Carlo Tree Search (AB-MCTS). This technology leverages insights from "The AI Scientist," an earlier Sakana AI project featured in the journal Nature for automating scientific discovery.

To illustrate, consider modern chess engines. They don't merely guess; they simulate thousands of potential future moves, assessing the strength of each resulting position before making a decision. Marlin's AB-MCTS engine applies a similar principle to research, treating the process as a branching tree of possibilities.

The genesis of this technology dates back to June 2025, when Sakana AI introduced the framework alongside its research paper, "Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search." To foster developer experimentation with collective AI intelligence, the company open-sourced the underlying algorithm as the TreeQuest software library under the Apache 2.0 license, laying the groundwork for what would become the proprietary Marlin product.

Multi-LLM Approach and Strategic Depth

Traditional methods for enhancing LLM reasoning often involve "repeated sampling," running a model multiple times in parallel and hoping for a correct answer. This approach is blind, lacking the ability to evaluate intermediate steps or adapt based on feedback. AB-MCTS replaces this with a principled, multi-turn approach driven by a Bayesian decision framework.

As Marlin constructs a strategy report, the system views the research as a branching tree of possibilities. At each node, the algorithm dynamically balances two behaviors based on external feedback signals: "Going Wider" (exploration), which involves spawning new hypotheses when current paths yield diminishing returns, and "Going Deeper" (exploitation), which methodically refines and builds upon promising solutions.

Crucially, this system extends into Multi-LLM AB-MCTS, introducing a third dimension: the ability to dynamically select which model to invoke for specific sub-tasks. This treats leading frontier models as a plug-and-play collective intelligence network. Sakana AI's architecture allows an orchestration model to delegate initial ideation to one LLM while a reasoning-heavy model audits, verifies, and corrects errors, ensuring reports are not merely long but the highly vetted product of systemic, automated trial-and-error.

Enterprise Focus: Licensing and Data Security

Sakana Marlin is a commercial Software-as-a-Service (SaaS) offering strictly for corporate entities, organizations, and sole proprietors, not a general consumer tool. Data handling is paramount for enterprises, and Marlin operates under a strict, enterprise-grade data policy. Neither Sakana AI nor its external AI service providers use customer data or inputs for model training or fine-tuning without explicit client consent. Even with consent, data undergoes heavy processing to remove personally identifiable information, providing vital security for sensitive corporate research.

The commercial licensing reflects its enterprise nature:

• Pay-as-you-go: 100 credits per run, add-on credits at ¥98 ($0.61 USD).
• Pro Plan: ¥150,000 ($935.68 USD) monthly for 2,000 credits; add-on credits at ¥90 ($0.56 USD).
• Team Plan: ¥400,000 ($2,495.14 USD) monthly for 6,000 credits; add-on credits at ¥85 ($0.53 USD).
• Enterprise: Custom quotes, dedicated support, and tailored credit allocations.

Sakana AI's Vision and Impact

Sakana AI's rapid transition into a commercial enterprise is bolstered by its founders' influential backgrounds. Co-founded in Tokyo in 2023 by Llion Jones, a co-author of Google's seminal "Attention Is All You Need" paper and the coiner of the term "transformer," and David Ha, a former Google Brain researcher, the startup deliberately chose to build outside the Silicon Valley ecosystem. Jones has vocalized his weariness with the industry's singular focus on scaling monolithic models, arguing it stifles innovation.

Sakana AI’s philosophy is rooted in biomimicry and evolutionary computing, inspired by collective intelligence observed in nature, such as schools of fish (reflected in its name, "sakana" being Japanese for fish). Instead of one massive, general-purpose model, Sakana focuses on networks of smaller, specialized models collaborating dynamically. This approach has already yielded success, including its ALE-Agent winning the AtCoder Heuristic Contest (AHC058) and its "RL Conductor" model orchestrating various frontier LLMs for state-of-the-art reasoning at lower costs.

The startup's trajectory has attracted significant investment, securing a Series B funding round by late 2025 that pushed its valuation past $2.6 billion. Its impressive roster of strategic investors includes Khosla Ventures, Lux Capital, NEA, Nvidia, Google, Mitsubishi UFJ Financial Group (MUFG), Citi, and Salesforce, signaling its potential to redefine corporate AI infrastructure. Early beta testers, including financial institutions and consulting firms, have praised Marlin for discovering unforeseen angles, matching human comprehensiveness without bias, and delivering high-quality, primary research.

FAQ

Q: What is Sakana Marlin and how does it differ from other AI tools?

A: Sakana Marlin is an autonomous, B2B AI research agent designed for "ultra deep research" and long-horizon reasoning. Unlike typical chatbots that provide instantaneous, surface-level answers, Marlin operates for up to eight hours, conducting extensive research to produce detailed, cited 100+ page strategy reports and executive slides for enterprise clients.

Q: What is the underlying technology that powers Marlin's long-horizon reasoning?

A: Marlin is powered by Sakana AI's Adaptive Branching Monte Carlo Tree Search (AB-MCTS) engine. This technology allows the AI to dynamically balance exploration (generating new hypotheses) and exploitation (refining existing solutions) within a "Multi-LLM AB-MCTS" framework, orchestrating multiple specialized AI models to collaborate on complex research tasks.

Q: Is Sakana Marlin suitable for individual use, and how does it handle data privacy?

A: No, Sakana Marlin is strictly a commercial software-as-a-service (SaaS) offering intended for corporate entities, organizations, and sole proprietors, not general consumer use. It adheres to a strict, enterprise-grade data policy, meaning customer data and inputs are not used for model training or fine-tuning without explicit client opt-in consent, and personally identifiable information is heavily processed to ensure privacy.