Sakana AI's First Commercial Product Takes Eight Hours to Answer — On Purpose

The Tokyo lab behind some of AI's most cited research is betting enterprise clients will pay nearly $1,000 a month for an agent that thinks slowly and delivers 100-page strategy reports.

Written by Theo Kline · Bureau Tech · Jun 18, 2026

The Slowest AI in the Room Is the Point

Sakana AI's first commercial product does something almost no AI product in 2026 is designed to do: it makes you wait. Marlin, billed as a "Virtual CSO" (Chief Strategy Officer), runs autonomous reasoning loops for up to eight hours before delivering a structured, cited strategy report that can exceed 100 pages. That is not a bug in the product roadmap. It is the product.

The Tokyo-based startup — co-founded by Llion Jones, a co-author of Google's 2017 "Attention Is All You Need" paper that introduced the transformer architecture now underpinning most modern AI — launched Marlin commercially this week with tiered pricing aimed squarely at corporations, financial institutions, and think tanks.

How It Actually Works

The workflow breaks from standard large language model (LLM) interactions. A user submits a research topic, answers a brief scoping exchange, then steps away. Marlin takes over: forming hypotheses, gathering web data, cross-referencing sources, and mapping causal relationships across a business environment — autonomously, for hours.

The engine powering this is Adaptive Branching Monte Carlo Tree Search (AB-MCTS), a framework Sakana first published and open-sourced in June 2025 under the Apache 2.0 license. The chess analogy is apt: rather than guessing at an answer, the system plays out thousands of research paths, evaluating each before committing. At every branch point, the algorithm chooses between two behaviors — spawning new hypotheses when a line of inquiry stalls (exploration), or drilling deeper into a promising thread (exploitation).

Sakana extends this into Multi-LLM AB-MCTS, where different frontier models are invoked for different sub-tasks. An orchestration model might delegate ideation to one LLM and route verification to a reasoning-heavy model. Sakana has not disclosed which external model providers it uses.

The Enterprise Pitch

For enterprise buyers, the data policy may matter as much as the output quality. Sakana says neither it nor its AI service providers will use customer inputs to train or fine-tune models without explicit opt-in consent — a meaningful commitment for clients running sensitive competitive or M&A research through the system.

Pricing is structured in three tiers: a pay-as-you-go option where a single Marlin run costs 100 credits (add-ons at ¥98/$0.61 each); a Pro Plan at ¥150,000 (~$936) per month for 2,000 credits; and a Team Plan at ¥400,000 (~$2,495) per month for 6,000 credits. Enterprise quotes are custom.

A closed beta involving roughly 300 professionals at financial institutions and consulting firms preceded the launch. Early feedback cited the tool's ability to surface angles human researchers missed and its reliance on primary rather than recycled secondary sources. That is encouraging, but beta cohorts are not retention data.

What the Funding Doesn't Prove

Sakana's Series B pushed its valuation past $2.6 billion, with backers including Nvidia, Google, Khosla Ventures, MUFG, Citi, and Salesforce. That is a credible cap table. It is not evidence that enterprises will pay a recurring four-figure monthly fee for a tool that takes a workday to return results, however good those results are.

The real test is whether Marlin fits into actual enterprise workflows — where procurement cycles are long, IT security reviews are thorough, and the bar for replacing a human strategy team is higher than any press release can clear. Sakana's research pedigree is genuine. The commercial durability of slow AI in a fast-AI market is still an open question.

Key takeaways

Marlin deliberately trades speed for depth — it runs for up to eight hours per query, positioning itself against the instant-answer paradigm that defines most AI tools.
The engine is built on Adaptive Branching Monte Carlo Tree Search (AB-MCTS), a technique Sakana open-sourced in mid-2025 that dynamically balances exploring new hypotheses versus refining promising ones.
Customer data is not used to train underlying models without explicit opt-in consent — a meaningful differentiator for enterprises handling sensitive M&A or competitive intelligence.
Sakana's $2.6 billion post-money valuation and investor roster (Nvidia, Google, MUFG, Citi, Salesforce) signals serious institutional confidence, but the product still needs to prove retention beyond its closed beta.
Co-founder Llion Jones, who coined the term 'transformer,' has publicly called the industry's fixation on scaling monolithic models a creativity trap — Marlin is the commercial argument for his alternative thesis.

FAQ

What is Sakana Marlin and who is it for?

Marlin is an autonomous B2B research agent that runs for up to eight hours to produce structured strategy reports exceeding 100 pages. It is designed for enterprises, financial institutions, and think tanks — not individual consumers.

What is AB-MCTS and why does it matter here?

Adaptive Branching Monte Carlo Tree Search (AB-MCTS) is a reasoning framework that treats research as a branching tree of possibilities, dynamically choosing between exploring new hypotheses and refining promising ones based on feedback signals. It replaces the 'repeated sampling' approach — running a model many times and hoping for a good answer — with a principled, multi-turn method. Sakana open-sourced the underlying algorithm in 2025 before building Marlin on top of it.

Will Sakana use my company's data to train its models?

According to Sakana's published data policy, neither Sakana nor its external AI service providers will use customer inputs for model training or fine-tuning without explicit opt-in consent from the client.

How much does Marlin cost?

A single run costs 100 credits on the pay-as-you-go tier, with add-on credits at ¥98 (~$0.61) each. The Pro Plan is ¥150,000 (~$936) per month for 2,000 credits. The Team Plan is ¥400,000 (~$2,495) per month for 6,000 credits. Enterprise pricing is negotiated separately.

Who founded Sakana AI and what is their background?

Sakana AI was founded in Tokyo in 2023 by Llion Jones — a co-author of Google's 2017 'Attention Is All You Need' paper and the person credited with coining the term 'transformer' — and David Ha, a former Google Brain researcher and ex-head of research at Stability AI.

Citations

When deep research isn't enough for your business: Sakana AI launches 'ultra deep research' agent for 100+ page reports in 8 hoursSakana AI launched Marlin, a B2B autonomous research agent that produces 100-plus-page strategy reports after up to eight hours of continuous reasoning.
Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree SearchSakana AI introduced and open-sourced the AB-MCTS framework in June 2025 under the Apache 2.0 license via a library called TreeQuest.
VentureBeat — Sakana AI Marlin coverageSakana AI's Series B pushed its post-money valuation past $2.6 billion, with investors including Nvidia, Google, MUFG, Citi, and Salesforce.
VentureBeat — Sakana AI Marlin coverageSakana AI ran a closed beta with approximately 300 professionals at financial institutions, consulting firms, and think tanks before the commercial launch.