Sakana launches Marlin as a Virtual CSO with up to 8-hour autonomous research
Sakana launched Marlin, a Virtual CSO that runs for up to 8 hours, forms hypotheses, browses sources, and returns slide decks plus reports. It turns Sakana’s long-horizon reasoning work into a shipped deep-research product.
TL;DR
- SakanaAILabs' English launch thread says Sakana Marlin is the company's first commercial product, positioned as a "Virtual CSO" for business research rather than a general chat assistant.
- According to hardmaru's launch post and SakanaAILabs' Japanese product thread, Marlin can run autonomously for up to roughly 8 hours, forming hypotheses, gathering sources, and verifying findings before returning slides plus a long report.
- SakanaAILabs' launch thread ties Marlin directly to Sakana's long-horizon reasoning work and AB-MCTS, while hardmaru's post calls it a productionized version of the team's core research.
- SakanaAILabs' Japanese product thread says the output can stretch to around 100 pages, and SakanaAILabs' English thread frames the target workflow as strategy work that might otherwise take a CSO and small team weeks.
- Availability is live now, with pay-per-use plus Pro, Team, and Enterprise plans, per SakanaAILabs' English launch thread and SakanaAILabs' Japanese release post.
SakanaAILabs' Japanese release post links straight to the launch materials, SakanaAILabs' English thread makes the product claim unusually explicit by calling it Sakana's first commercial product, and hardmaru's post spells out the research lineage, including AB-MCTS and The AI Scientist. The more interesting operational detail sits in the Japanese product thread, which says Marlin can run for about 8 hours and return structured slides plus a report of up to roughly 100 pages.
Virtual CSO
Sakana is shipping Marlin as a business research agent with a very specific buyer story. The pitch is not generic deep research, it is a virtual chief strategy officer that takes a single research topic and runs with it.
The company says that single prompt is the only required user input. From there, Marlin handles hypothesis formation, source gathering, and self-verification before producing decision-ready deliverables.
8-hour research loop
The core product claim is duration. While most deep-research products optimize for minutes, SakanaAILabs' English thread and hardmaru's post both center an autonomous run that can last up to roughly 8 hours.
Sakana describes the loop in four steps:
- form hypotheses
- gather information
- verify findings
- resolve contradictions
That last step matters because it pushes Marlin past summarization language and into an agentic research workflow. hardmaru's launch post says the system navigates the web and resolves contradictions during the run, while the Japanese thread says it keeps iterating through issues buried in large information sets.
AB-MCTS in the product
The technical hook is AB-MCTS, the multi-model reasoning method Sakana has been publishing on. SakanaAILabs' English launch thread says Marlin is built on two years of long-horizon reasoning work and uses AB-MCTS to coordinate multiple models more effectively.
Hardmaru goes further than the corporate thread and says Marlin is the direct productionization of Sakana's core breakthroughs, naming both AB-MCTS and The AI Scientist. That makes this launch less like a wrapper around existing research tools and more like Sakana's first attempt to turn its inference-time compute story into a paid product.
Slides and 100-page reports
Sakana is selling output format as much as model behavior. According to SakanaAILabs' Japanese product thread, Marlin returns structured summary slides and a research report that can reach roughly 100 pages.
The target job is strategy work, market research, risk analysis, and competitive analysis. SakanaAILabs' English thread says the system is designed for the kind of investigation a CSO and a small team might otherwise spend weeks on.
Beta footprint and pricing
The launch materials add two concrete go-to-market details that are easy to miss in the tweets alone. The linked release summary attached to SakanaAILabs' Japanese announcement says Marlin was refined through a closed beta that started in April 2026, with around 300 professionals across finance, operating companies, consulting, and think tanks.
That same material says Marlin is self-serve from day one, with a pay-per-use option that carries no monthly fee, plus Pro, Team, and Enterprise tiers. SakanaAILabs' English launch thread also says Marlin is the first of many products Sakana plans to ship, which makes this both a product launch and a marker that the company is moving from research branding into a commercial lineup.