Sakana AI bets against the 'strongest AI' era: orchestration, shipped fast

The dominant play in AI today, as Sakana AI's head of product development sees it, is polishing one supreme model and handing it to the whole world. Sakana is running the other way. Since March 2026, the Tokyo firm has shipped four products in quick succession: Sakana Chat, Sakana Marlin, Sakana Fugu, and Sakana Translate. Sota Omura says the pace is deliberate, and the strategy underneath it is a bet against the 'strongest AI' era.

Ship first, correct fast

Sakana Chat arrived in March 2026, with Marlin, Fugu, and Translate following in short order. Omura credits a firm directive from CEO David: this would be the year Sakana ships products. But he argues the pace is becoming standard practice in the AI era, even though it departs sharply from the conventional approach. Prototypes that once took months are finished in weeks, sometimes days. It is faster to 'ship first and learn from real reactions' than to spend a year investigating before building. The four products are deliberately not built for one industry. Translation, research, and chat-based consulting show up in daily work across most industries. Specialized versions of Marlin for finance and manufacturing are already under active exploration.

The case against the single best model

The products split into two groups. Chat and Translate let people experience Namazu, the company's flagship model, with additional training that tunes it closely to the Japanese context. Fugu and Marlin explore orchestration, the practice of making several models work in concert. Fugu, offered through an API, handles tasks with higher performance than any single model alone; Marlin, built on the AB-MCTS technique for deep reasoning, is meant to grow into a virtual chief strategy officer. Together they are the commercial expression of a thesis seventnews has called making the model choice the router's problem.

There is a warning built into the design. A world where everyone leans on the same 'strongest AI' is a world to avoid. Access to quality models already hinges on money, nationality, and geography, and supply chain risk is hard to ignore. The pull toward a single champion is easy to see in the Chatbot Arena, where nearly five million votes now rank models by blind human preference, per the leaderboard's ranking data. Sakana's answer is diversity: the ways people work with AI are healthier when they stay varied, just as ecosystems are.

The thesis extends beyond the products. Sakana argues Japan can secure sovereignty in post-training, adapting open models to local language and values, rather than outspending U.S. and Chinese tech giants on pre-training. Namazu is the demonstration. Leadership has said Sakana Fugu can match Mythos in certain domains because orchestration compensates for individual model weakness, while conceding U.S. companies still have better single-model performance. The benchmarks were not disclosed. In its Series B round, Sakana AI accepted investment from In-Q-Tel, the venture capital arm founded by the CIA, a move the company tied to growing interest in defense applications. Sakana states publicly that its products will not take autonomous lethal action, in line with Japan's constitutional framework.

Omura describes the goal in concrete terms:

"An AI sat in on the board meeting and, through discussion with the human directors, helped decide where to invest. How many companies will be able to say that in the years ahead?"

From research tool to decision platform

Marlin shows where Sakana wants to go. Omura's ambition is a platform for decision-making, built on 'hypothesis thinking,' the method proposed by Kazunari Uchida, who headed the Boston Consulting Group in Japan: investigate, hypothesize, act, verify, revise. Today Marlin covers only the first step. Omura describes an investment decision as a spiral: research yields a list of 100 to 200 candidates, the ten most promising get studied in depth, scenarios get simulated, and the cycle repeats. Consulting has held its ground despite AI, he notes. That says something about how uncertain and ambiguous decision making really is. Most AI use stops at piecemeal efficiency, writing Excel formulas and arranging schedules. The gap shows up in a 957,253-record benchmark corpus, where agents surge in coding but stall where enterprises need them, per the Messier benchmark analysis. Omura's goal is humans and AI bringing hypotheses and evidence to the same table.

An organization built on overlap

The organization is built to survive the expansion. Omura favors business units, where every function chases the same numbers, over a functional structure split by specialty. He concedes the cost: overlapping roles demand that everyone commit to the whole process, a professional skill rather than good intentions. His references are Ikujiro Nonaka and Hirotaka Takeuchi, whose knowledge-creation theory holds that new knowledge emerges when people with different experiences put tacit knowledge into words. Their 1986 Harvard Business Review article, 'The New New Product Development Game,' later fed into Scrum, the agile method now standard in software. He also cites the 'diversity trumps ability' theorem of Lu Hong and Scott Page. The product team has about 15 people, with marketing and sales still to be built.

The honest version of the bet is that the gap remains open. Sakana AI concedes U.S. labs hold the edge on single-model performance, and the marquee claims for Fugu rest on evaluations the company has not published. The distance between benchmark claims and deployment is familiar: on SWE-bench Verified, top models hit 96%, but on private enterprise code they barely clear 23%, per the benchmark comparison. The company is testing whether orchestration closes that gap fast enough to make the 'strongest AI' race beside the point. If Omura is right, the next year will show it.