Tokyo-based AI innovator Sakana AI has unveiled its inaugural commercial offering, Sakana Marlin, positioning it as a pivotal advancement in enterprise AI solutions.
Marketed as a “Virtual Chief Strategy Officer” (CSO), Marlin operates as an autonomous, business-to-business research agent. Crucially, it diverges from the rapid, instantaneous text generation characteristic of contemporary chatbots, prioritizing instead a methodology of deep, long-horizon strategic analysis.
What distinguishes Marlin within the current AI landscape is its temporal approach. Rather than yielding answers in seconds, it engages in continuous, self-directed reasoning cycles lasting up to eight hours. This extended processing allows for the generation of comprehensive, meticulously cited strategy reports and executive slide decks, often reaching 100 pages in length. The company provides sample reports generated by Marlin on its product website.
Marlin is available immediately through Sakana AI’s website, with a flexible pay-as-you-go pricing structure. The platform is exclusively designed for enterprise applications, targeting corporations, financial institutions, and think tanks that require in-depth strategic intelligence.
The generative AI sector has largely been defined by speed. For the past two years, the industry’s benchmark has been the ability to produce content—poems, code, or superficial summaries—in milliseconds. However, the enterprise frontier is rapidly evolving from shallow, rapid output to profound, methodical analysis.
With Sakana Marlin, major enterprises are shifting their focus from how quickly an AI can respond to how deeply it can reason.
The Product: A Virtual CSO
Deploying Sakana Marlin grants businesses a fundamentally different operational workflow compared to typical interactions with large language models (LLMs). Instead of engaging in iterative prompt engineering, users provide a core research topic. Following a brief initial dialogue to refine the scope, the human element steps back entirely.
For the subsequent hours, Marlin functions as a self-sufficient digital strategy department. It independently formulates hypotheses, conducts web-based data collection, cross-references sources for validation, and maps complex causal dynamics within business environments. Its objective is to discern the optimal strategic path amidst vast amounts of information.
Consider it less a search engine and more akin to a junior strategy consultant tasked with a complex problem, equipped with a whiteboard and internet access. The strategic prompt is provided in the morning, and by day’s end, the system delivers a polished, professional-grade strategic analysis.
The final output from Marlin is not a generic block of text; rather, it is a structured compilation of strategic options, complete with executive summaries, appendices, detailed references, and a deeply researched report.
Sakana AI has showcased Marlin’s capacity for complex synthesis through real-world scenario analyses, including detailed resolution pathways for a theoretical blockade of the Strait of Hormuz, mapping the intricate global landscape of AI regulation, and analyzing macroeconomic shifts such as the resurgence of “bond vigilantes.”
The company indicates that Marlin utilizes a combination of AI models but has not specified names or providers. Inquiries have been made for further details.
VB Transform · July 14–15 · Menlo Park · LLMs, ops & evals
Standard benchmarks fail. Amazon and Waymo explain what they test instead.
The evals track goes deep on the four dimensions of reliability — consistency, robustness, predictability, safety — and how teams at Amazon and Waymo are operationalizing them in production.
See the full agenda →
The Engine of Long-Horizon Reasoning
Marlin represents the commercial realization of Sakana AI’s extensive research and development over the past two years.
The product’s engine is built upon Sakana AI’s proprietary Adaptive Branching Monte Carlo Tree Search (AB-MCTS) technology. This framework is derived from “The AI Scientist,” a prior research project that successfully automated aspects of the scientific discovery process, from ideation to peer review, as documented in the journal Nature.
To illustrate the mechanism, consider the analogy of modern chess engines. When a computer plays chess, it evaluates thousands of potential moves and resulting board states before committing to an action. Marlin’s AB-MCTS engine applies a similar principle to its research processes.
Inside the Engine: The Mechanics of AB-MCTS
The technological lineage of AB-MCTS dates back to June 2025, when Sakana AI first presented the framework publicly alongside their research paper, “Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search.”
To foster experimentation in collective AI intelligence, the company subsequently released the underlying algorithm as an open-source library named TreeQuest, under the permissive Apache 2.0 license. This open-source initiative provided the foundational technology for the evolution into the commercial Marlin product a year later.
Traditionally, enhancing the reasoning quality of large language models often involves “repeated sampling”—running the model multiple times in parallel and selecting the best outcome. However, this brute-force approach lacks self-assessment capabilities and cannot adapt based on intermediate results.
AB-MCTS introduces a structured, multi-turn approach guided by a Bayesian decision framework. As the AI constructs a strategy report, the research process is modeled as a branching tree of possibilities. At each node, the algorithm dynamically balances two core functions, informed by external feedback signals:
-
Exploration (Going Wider): Initiating new, alternative hypotheses or potential responses when the current research path shows diminishing returns or encounters inconsistencies.
-
Exploitation (Going Deeper): Methodically refining, validating, and expanding upon existing candidate solutions that demonstrate significant strategic potential.
The transformation from a research concept to a commercial engine is achieved through its extension into Multi-LLM AB-MCTS.
Sakana AI’s architecture introduces a third dimension to the search tree: the dynamic selection of which AI model to employ for specific sub-tasks. This approach treats leading industry models as components within a collaborative intelligence network.
According to the company’s technical documentation, the engine can orchestrate highly diverse models. It enables an orchestration model to delegate initial ideation to one LLM while leveraging a distinct reasoning-focused model to audit and correct errors generated earlier in the process.
By optimizing compute resources during inference—harnessing the unique capabilities of multiple foundation models over thousands of automated cycles—AB-MCTS provides the necessary analytical rigor for Marlin. This ensures that the extensive strategy reports are not merely lengthy AI outputs but the rigorously vetted results of systematic, automated problem-solving.
Licensing, Data, and Enterprise Implications
It is important to emphasize that Sakana Marlin is strictly a commercial Software-as-a-Service (SaaS) offering intended for enterprise clients, organizations, and independent professionals, not for general consumer use.
For enterprises, licensing terms and data handling protocols are critical considerations for adoption. Unlike many consumer-grade AI tools that may utilize user inputs for model training, Sakana Marlin adheres to a stringent enterprise-grade data policy.
Sakana AI and its service providers will not use customer data or inputs for model training or fine-tuning without explicit opt-in consent from the client. Even with consent, data is anonymized to remove personally identifiable information, ensuring robust security for sensitive research such as M&A activities, product strategies, or proprietary market analyses.
The commercial licensing is structured into tiered pricing models tailored for enterprise needs:
-
Pay-as-you-go: Users purchase credits on demand, with each research run costing 100 credits. Additional credits are priced at ¥98 (approximately $0.61 USD) each.
-
Pro Plan: Priced at ¥150,000 (approximately $935.68 USD) per month, this tier includes 2,000 credits, reducing the cost of add-on credits to ¥90 ($0.56 USD).
-
Team Plan: Designed for larger departments, this plan costs ¥400,000 (approximately $2,495.14 USD) per month and provides 6,000 credits, with add-on credits at ¥85 ($0.53 USD) each.
-
Enterprise: Custom quotes are available, offering dedicated support and flexible credit allocations.
Why Sakana AI is a Company to Watch
Sakana AI’s emergence as a significant player in the enterprise AI market is underpinned by the distinguished backgrounds of its founders, who played key roles in the foundational developments of the current generative AI revolution.
Founded in Tokyo in 2023, the startup was co-established by Llion Jones, a co-author of the pivotal 2017 paper “Attention Is All You Need” who also coined the term “transformer,” and David Ha, a former Google Brain researcher and ex-head of research at Stability AI.
The decision to establish Sakana AI outside the traditional Silicon Valley hub was a deliberate move to foster innovation distinct from the prevailing AI ecosystem. Jones has articulated concerns that intense investor pressure and an overemphasis on scaling single, monolithic models have stifled creativity and potentially overlooked emergent breakthroughs.
To counter this perceived stagnation, Jones and Ha have structured Sakana AI around principles of biomimicry and evolutionary computing. The company’s name, referencing Japanese fish, reflects its core technical philosophy: leveraging collective intelligence akin to schools of fish or ant colonies. Instead of focusing on developing one massive, all-encompassing foundation model, Sakana AI’s research emphasizes deploying networks of smaller, specialized models that collaborate dynamically.
This approach suggests that by treating individual AI models as part of a “dream team” with complementary strengths, systems can achieve more robust and cost-effective reasoning than relying solely on sheer scale.
This nature-inspired methodology has already demonstrated success in competitive evaluations. Sakana AI has made significant advancements in “inference-time scaling”—optimizing computational resources during problem-solving to enable models to iterate and refine solutions over extended periods. In early 2026, Sakana’s ALE-Agent secured first place in the complex AtCoder Heuristic Contest (AHC058), outperforming numerous competitors by autonomously developing and testing hundreds of solutions within a four-hour timeframe.
Additionally, Sakana introduced “RL Conductor,” a compact 7-billion-parameter model trained via reinforcement learning to orchestrate tasks among diverse worker models. This system has achieved state-of-the-art performance on reasoning benchmarks at a significantly reduced computational cost.
Sakana AI’s rapid progression from a research-focused lab to a commercial software provider has garnered substantial attention from major global investors. By late 2025, the company secured a significant Series B funding round, valuing it at over $2.6 billion. Key investors include Khosla Ventures, Lux Capital, NEA, Nvidia, and Google, alongside strategic partners like Mitsubishi UFJ Financial Group (MUFG), Citi, and Salesforce, indicating Sakana’s potential to redefine corporate AI infrastructure.
Community Reactions and Field Testing
Sakana AI’s development of long-horizon, autonomous agents was preceded by a rigorous closed beta test initiated in April 2026. Approximately 300 professionals from financial institutions, consulting firms, and think tanks participated, providing feedback that highlighted a clear qualitative difference between standard generative chatbots and Marlin’s autonomous, evidence-based approach.
Participants noted Marlin’s ability to uncover novel strategic perspectives and provide comprehensive analyses free from human bias. A senior consultant at a major Tokyo firm remarked that the tool “exceeded expectations by discovering angles we hadn’t even imagined,” while a cybersecurity division praised the system for delivering “a highly convincing report driven by high-quality, primary research.”
The announcement has resonated within the broader tech community, reflecting a growing demand for sophisticated autonomous agents. As the AI industry matures, the value proposition is shifting from rapid content generation to advanced analytical capabilities.
Sakana Marlin facilitates this shift by delegating the intensive task of mapping complex dynamics to an agent capable of sustained reasoning, thereby freeing human executives to focus on strategic decision-making and execution.
Business Style Takeaway: Sakana AI’s Marlin represents a significant strategic pivot in enterprise AI, moving beyond rapid response to deep, autonomous strategic analysis. This signals a market evolution where businesses are prioritizing sophisticated reasoning capabilities for complex problem-solving over immediate output, indicating a future focus on AI agents that augment executive decision-making rather than merely automate tasks.
Based on materials from : venturebeat.com