Sakana Marlin is the product Sakana AI shipped before Fugu — its first commercial product, launched June 15, 2026 — and it’s a different animal entirely: not a model you call per token, but an autonomous strategy-research agent that runs for hours and comes back with a report. Sakana markets it as “Your Virtual CSO.” This guide covers what it does, how the credit pricing works (in yen, as published), how it relates to Fugu, and who it’s actually for. Verified against Sakana’s official product page and launch announcement on August 19, 2026.
What Sakana Marlin is
Sakana’s own description sets the frame: Marlin is “not just another deep research assistant” but an autonomous agent designed to take on “the kind of substantial strategy research that a Chief Strategy Officer and a small team might otherwise spend weeks on.” You set a research topic; it works without further human input — repeatedly forming hypotheses, gathering information, navigating the web, and resolving contradictions — through up to eight hours of continuous, autonomous reasoning.
The output claim is equally specific: not summaries, but mapped causal relationships organized into structured strategic options — “exhaustive strategy reports” meant to leave the human with the decision itself. Under the hood, Sakana’s FAQ credits a technique it calls AB-MCTS, which drives the autonomous hypothesis → exploration → verification loop; the deliberate design choice is long-horizon, sample-efficient reasoning rather than fast text generation. A single research task takes roughly eight hours as a guideline, varying with topic complexity — and per the FAQ, you can cancel a run in progress, but the credits are still consumed.
Pricing: credits, in yen
Marlin prices in Japanese yen, on a credit system where every research run costs 100 credits. The published tiers (verbatim from the official page, August 19, 2026; card signup, same-day start, and a free-trial application both offered):
| Plan | Price | Credits | Add-on credits |
|---|---|---|---|
| Pay as you go | Per-use | Buy on demand | ¥98 / credit |
| Pro | ¥150,000 / month | 2,000 / month | ¥90 / credit |
| Team | ¥400,000 / month | 6,000 / month | ¥85 / credit |
| Enterprise | Custom quote | Customized | Discounted; dedicated support |
Do the run math: at pay-as-you-go rates, one 100-credit run is ¥9,800 (approx. $60–65 US, as of Aug 2026 — our conversion, not an official price); on Pro, the monthly fee works out to 20 runs (¥7,500 per run, approx. $47); on Team, 60 runs (¥6,667 per run, approx. $42). In other words, a Marlin run is priced like a deliverable, not a chat session — which is exactly the comparison Sakana invites by framing it against weeks of strategy-team work.
Marlin and Fugu: sibling products, different species
It’s easy to assume Marlin is “Fugu with a report template.” The official materials don’t say that, and the products differ in kind. Fugu is infrastructure: a model API you build on, billed per token, orchestrating a pool of frontier models per request. Marlin is a finished application: a research service billed per run, no API surface advertised, sold to strategy and business-intelligence users rather than developers. What they share is the company’s core conviction — spending inference-time compute deliberately (Fugu through multi-agent scaffolds, Marlin through hours-long search loops) beats generating fast answers.
Practical rule of thumb: if you’d brief a consultant, that’s Marlin-shaped work; if you’d assign an engineer or wire an API, that’s Fugu-shaped.
Who it’s for — and who it isn’t
The official positioning targets strategy work: competitive landscapes, market entry, the research that precedes decisions — with humans kept for “the highest-value work of all: the decisions themselves.” The Japanese-market orientation is also plain: a yen-priced product from a Tokyo company whose enterprise business runs deep in Japanese finance (see our company profile). It is not for interactive research sessions (runs take hours and consume credits even if cancelled), not a per-token API you can meter finely, and not — as of August 2026 — priced casually: the entry point to subscription value is ¥150,000 a month.
One honest unknown, flagged per our standards: Sakana’s public page doesn’t state which models power Marlin’s reasoning, and we won’t guess. What’s published is the technique framing (AB-MCTS, inference-time scaling) and the product behavior.
What does Sakana Marlin actually deliver?
A structured strategy report: per Sakana, it maps causal relationships in complex business environments and organizes them into strategic options, after up to eight hours of autonomous hypothesis-forming, web research, and contradiction-resolving on the topic you set.
How much does one Marlin run cost?
Every run costs 100 credits. Pay-as-you-go credits are ¥98 each (so ¥9,800 per run — approx. $60–65 US as of Aug 2026, our conversion); the Pro plan (¥150,000/month, 2,000 credits) works out to ¥7,500 per run; Team (¥400,000/month, 6,000 credits) to ¥6,667. Yen prices are the official ones.
Can I stop a run partway?
Yes — Sakana’s FAQ says a run can be cancelled at any time, but the credits are consumed regardless. Treat each run as a committed purchase.
Is Marlin built on Fugu?
Sakana doesn’t say which models power Marlin, and we don’t speculate. Officially, Marlin is characterized by its AB-MCTS-driven autonomous research loop and long-horizon inference-time reasoning; it launched a week before Fugu as the company’s first commercial product.
Is there a free trial?
The official page offers a free-trial application alongside card-signup for immediate paid use (as of August 19, 2026). Terms of the trial aren’t published on the page; apply through sakana.ai/marlin.
The rest of the lineup: Sakana AI company profile, What is Sakana Fugu?, and the Namazu guide. Building rather than briefing? Start with the Fugu API guide.