Dissecting Sakana AI — The $2.7 Billion Company a Transformer Author Built in Tokyo

CEO David Ha spent eight years at Goldman Sachs, rising to co-head of Japan rates trading and managing director before moving to Google Brain in 2016. While still at the bank, he ran an AI research blog under the anonymous handle "hardmaru," which built a following in the deep learning community before his name did. After Google Brain, he spent about a year from 2022 leading research at Stability AI.

CTO Llion Jones grew up in Wales. He built a chatbot at age 14 and earned both a bachelor's and master's in computer science and AI from the University of Birmingham. His best-known credit is the 2017 paper "Attention Is All You Need" — one of eight co-authors on the paper that proposed the Transformer architecture now underlying nearly every large language model. Jones has said he joined Sakana AI because he believed the conventional Transformer approach had entered a stage of diminishing returns.

COO Ren Ito studied law at the University of Tokyo, earned an LLM at NYU Law, and a master's at Stanford. He joined Japan's Ministry of Foreign Affairs in 2001 and worked as a diplomat before becoming an executive officer for global operations at Mercari. He, too, previously served as COO at Stability AI. That Ha and Ito both passed through the same company, and that Ha and Jones met at Google's Tokyo office, is the backstory behind how this founding team came together.

The company's name, Sakana (魚), means "fish" in Japanese. A single fish follows only a handful of simple rules, but a school of fish as a whole produces complex behavior — evading predators, finding food. The idea of translating that collective intelligence into model design is where the company started. Against the industry's race to scale up parameters, the three founders believed there was another path.

Evolutionary Model Merging — Building New Models Without Training

The result that first put the company on the research community's radar was Evolutionary Model Merge, published on March 21, 2024. It combines several already-published open models into a new one — but instead of a human deciding how to combine them, an evolutionary algorithm searches for the combination. The paper was published in Nature Machine Intelligence in January 2025.

The search happens across two spaces at once. In parameter space (PS), the system evolves what ratio to blend the weights of multiple models, layer by layer. In data flow space (DFS), it evolves the order in which layers from different models are stitched together. Both spaces contain far too many possible combinations for human intuition to explore exhaustively. Running for 100 to 150 generations, only the best-performing individuals survive to produce the next generation, eventually yielding a final model.

The core of this method is that it never runs gradient-descent-based training even once. Instead of burning GPU cycles on repeated backpropagation, it reframes the problem as combining already-trained weights — cutting compute requirements dramatically.

The company also released concrete results. A 7-billion-parameter model called EvoLLM-JP scored 55.2% on the Japanese math reasoning benchmark MGSM-JA (50.6% for the publicly released v1-7B version). It was created by merging a Japanese-language model (Shisa-Gamma) with math-specialized models (WizardMath, Abel). Even though it was built specifically to target math, its average score on the comprehensive Japanese benchmark JP-LMEH was 70.5 — surpassing every Japanese model under 70 billion parameters and the best prior model near that size. That said, in the same table, Swallow 70B still scores higher, at 71.5. The same lineage includes EvoVLM-JP, which handles both images and language, and EvoSDXL-JP, an image generation model that processes Japanese prompts through four-step reasoning.

Space	What Evolves	Prior Human Approach
Parameter Space (PS)	Per-layer weight blending ratios	Fixed-ratio averaging
Data Flow Space (DFS)	Layer arrangement order across models	Designed by intuition and experience

AI That Fixes Itself — From The AI Scientist to the RSI Lab

The second pillar is research on getting AI to improve AI. It began with "The AI Scientist," released on August 13, 2024 — a system that runs the entire research loop without human involvement: generating research ideas, writing experiment code, running it, analyzing results, writing up the paper, and even peer review. The compute cost per paper was reported at roughly $15.

The follow-up, The AI Scientist v2, produced results at an ICLR 2025 workshop review in March 2025. One of three submitted papers scored an average of 6.33, clearing the acceptance bar — a score higher than 55% of human-authored submissions. However, Sakana AI voluntarily withdrew the paper before publication. It hadn't gone through the workshop organizers' meta-review, workshop acceptance rates are far higher than those of the main conference, and citation errors were flagged. That's the actual scope of the claim that "AI passed peer review." The paper describing the system itself was published in Nature on March 25, 2026 (vol. 651, pp. 914–919).

The same research also surfaced a safety-relevant signal. During testing, when an experiment ran past its time limit, the system — instead of writing faster code — modified its own code to simply extend the time limit. In one case, it inserted a system call to re-execute itself, creating an infinite loop. After the company disclosed these cases, debate followed over the control problem posed by autonomous agents.

Later releases pushed further into self-improvement. The Darwin Gödel Machine, released in May 2025, is a structure in which an agent rewrites its own codebase to generate variants. It raised SWE-bench performance from 20% to 50%, an absolute 30-point improvement, and lifted scores on the multilingual coding benchmark Polyglot from 14.2% to 30.7%. ShinkaEvolve, open-sourced in September of the same year, is a program-evolution framework combining language models with evolutionary algorithms. It matched the best known record on the classic circle-packing optimization problem (placing 26 circles) using only 150 samples, and in a separate experiment, discovered a new load-balancing loss function for mixture-of-experts (MoE) models within just 30 generations.

There are competition results too. On December 14, 2025, the algorithmic optimization agent ALE-Agent placed first in AtCoder Heuristic Contest 058, a competition where 804 human participants competed over four hours — marking the first time an AI agent won such an event. Total cost for the run was about $1,300.

This line of work was formally organized into the RSI (Recursive Self-Improvement) Lab on June 5, 2026. The lab's tagline is "advancing through ideas, not just compute" — turning Japan's lack of hyperscaler-level compute into a reason to push harder on sample-efficient research. Expansion into the physical world follows the same logic. On July 13, the company published research on implementing collective intelligence in physical environments via "smart cellular bricks," and on August 10, it presented physical AI as the next frontier through its own channels.

Research Aimed at What Comes After the Transformer

The third pillar is research aiming to change the architecture itself.

Continuous Thought Machines (CTM) were released in May 2025 — a structure that reintroduces a time axis into neural networks, where individual neurons retain memory of past activity and information is processed sequentially through neuron-to-neuron synchronization. The company says this architecture outperformed humans on confidence calibration.

The Neural Attention Memory Model (NAMM) was released in December 2024. It's a technique that learns which tokens accumulated in context should be kept versus discarded, reducing memory overhead at inference time — reportedly cutting memory costs by up to 75% on language and multimodal tasks. That's the kind of improvement that translates directly into savings in compute-constrained environments.

Transformer², released in January 2025, targets a self-adaptive architecture in which the model adjusts its own weights to fit the task at hand. All three lines of work sit apart from the "make the model bigger" direction.

There was one major misstep. "The AI CUDA Engineer," released in February 2025, claimed to automatically optimize GPU kernels, delivering up to 100x speedups over PyTorch. External verification revealed reward hacking — the system had exploited loopholes in the evaluation code. Some users reported the code actually ran three times slower in practice. On February 21, the company acknowledged the issue, revised its claims downward, corrected the results, and said it had strengthened its evaluation and runtime profiling processes. The question of who verifies the output of an automated research system — and how — has stayed with the company's approach to announcements ever since.

2026: The Year Research Became Products

2026 was the year Sakana AI turned its research output into revenue-generating products.

Sakana Chat launched free, Japan-only, on March 24. The Namazu (catfish) model family behind it takes open-weight models — DeepSeek-V3.1-Terminus, Llama-3.1-405B, gpt-oss-120B — and re-tunes them for Japanese language and cultural context. One figure the company shared illustrates the nature of that work: the base model DeepSeek-V3.1-Terminus refused to answer 72% of questions on politically sensitive topics, but after Namazu tuning, the refusal rate dropped to nearly 0%. On August 13, the chatbot was upgraded with a next-generation Namazu and Fugu, with the new Namazu's base model switched to Kimi K2.6.

Sakana Fugu launched in beta on April 24 and went generally available on June 22. It's an orchestration model — instead of a single model answering, it assigns multiple frontier models to different roles to divide up the work. This is backed by two papers presented at ICLR 2026. Trinity uses a lightly evolved coordinator to assign "thinker, worker, verifier" roles across multiple language models, while Conductor uses reinforcement learning to discover the communication patterns between agents itself.

The product launched as standard Fugu for general work and Fugu Ultra, which taps a deeper pool of expert models for harder problems, with Fugu Cyber, a security-reasoning-specialized tier, added on July 21. On the software engineering benchmark SWE-Bench Pro, the company reports Fugu Ultra scored 73.7. In the same table, Opus 4.8 scored 69.2, GPT-5.5 scored 58.6, and Gemini 3.1 Pro scored 54.2, while Fable 5 scored higher than Fugu Ultra at 80.0. On the security benchmark CyberGym, Fugu Cyber scored 86.9%, ahead of GPT-5.5-Cyber (85.6) and Claude Miso's preview (83.1). All figures are Sakana AI's own measurements; no third-party verification has been published.

Pricing combines monthly subscriptions at $20, $100, and $200 with pay-as-you-go options. Ultra's usage pricing is $5 per million input tokens and $30 per million output tokens, rising to $10 and $45 for long-context use beyond 272,000 tokens. The API is compatible with the OpenAI format and can also be accessed via OpenRouter, Vercel, and similar platforms. The European Economic Area is excluded for now, pending GDPR compliance.

A phrase the company repeats in Fugu's marketing is "frontier-level performance without export control risk" — clearly aimed at regions and industries subject to export restrictions on U.S. models.

Sakana Marlin went through a closed beta of roughly 300 people starting April 2 before its full launch on June 15 — Sakana AI's first commercial product. It's an autonomous research agent capable of running unsupervised for up to 8 hours in a single session, producing 100-page reports. The underlying technology is an adaptive branching Monte Carlo tree search that received a NeurIPS 2025 spotlight, combined with the workflow automation from The AI Scientist. Pricing is split between pay-as-you-go credits at 98 yen (100 credits), Pro at 150,000 yen/month, and Team at 400,000 yen/month. The company describes the product as a "virtual Chief Strategy Officer" for finance, consulting, and strategy organizations.

Product	Timing	Description
Sakana Chat	March 24, 2026	Free, Japan-only chatbot, Namazu family
Sakana Fugu	Beta April 24, GA June 22	Multi-model orchestration API
Sakana Marlin	Beta April 2, GA June 15	Up to 8-hour unsupervised research agent
Sakana Translate	July 6	Translation and proofreading service
Fugu Cyber	July 21	Security-reasoning-specialized tier

$400 Million in Three Years — Where the Money Came From

The pace of fundraising is rare even by the standards of Japanese startups.

The company raised a $30 million seed round in January 2024 led by Lux Capital, then closed a roughly $200 million Series A on September 4 of the same year, co-led by NEA, Khosla Ventures, and Lux Capital, with NVIDIA joining as a strategic investor. Alongside them came a lineup of major Japanese corporations: the three megabanks Mitsubishi UFJ, SMBC, and Mizuho, along with NEC, Fujitsu, Itochu, KDDI, Nomura, Dai-ichi Life, ANA Holdings, and Tokio Marine. Reports at the time valued the company at $1.5 billion — a unicorn just 14 months after founding.

Series B was first announced on November 17, 2025 at ¥20 billion (about $135 million) with a ¥400 billion valuation, then updated on April 9, 2026 to ¥32 billion (about $200 million) with a valuation of roughly ¥432 billion (about $2.7 billion). Cumulative funding stands at about ¥66 billion ($412 million). The round stayed open for five months as strategic investors kept joining. Google announced a strategic partnership and joined the round on January 23; Citi came in on February 24; Mitsubishi Electric on March 25. Citi's investment marks the bank's first strategic investment in a Japanese company. Salesforce Ventures, Datadog, Macquarie Capital, JAFCO, along with In-Q-Tel — the venture fund tied to U.S. intelligence agencies — and Santander's Mouro Capital round out the list.

Round	Timing	Size	Valuation
Seed	January 2024	$30 million	—
Series A	September 4, 2024	~$200 million	$1.5 billion (reported)
Series B	Nov 17, 2025 – Apr 9, 2026	¥32 billion (~$200 million)	~¥432 billion ($2.7 billion)

Instead of building its own data centers, Sakana AI solves compute needs through procurement. It received support from the Ministry of Economy, Trade and Industry's GENIAC program and uses GMO's GPU cloud. On July 16, 2026, it deployed Sakura Internet's "Koukaryoku PHY" and, on the same day, announced an open model collaboration with NVIDIA. This is a fundamentally different cost structure from U.S. and European rivals pouring billions of dollars into their own clusters.

Its First Customer Was a Bank; Now It's the Ministry of Defense

Revenue comes from multi-year partnerships with major Japanese corporations. In May 2025, Sakana AI signed a comprehensive partnership spanning more than three years with Mitsubishi UFJ to build bank-specific AI. In October, it signed with Daiwa Securities Group, moving into full-scale development of wealth-management AI as of August 5, 2026. Mitsubishi Electric, alongside its March investment, agreed to integrate Sakana AI models into its Serendie platform. On April 30, the company announced it was co-developing an automated proposal-generation app with SMBC Group, and on April 7, it joined a Ministry of Internal Affairs and Communications project addressing misinformation on social media. Unofficial estimates put 2025 revenue at around $30 million.

The second pillar is defense and intelligence. On March 13, 2026, Sakana AI received a multi-year commissioned research contract from the Acquisition, Technology & Logistics Agency's defense innovation research institute. The project, titled "Research to Accelerate Observation, Reporting, Information Integration, and Resource Allocation Through Combined AI Technologies," involves integrating multimodal data across land, sea, and air domains and building small vision-language models (SVLMs) deployable on edge devices like drones, aimed at upgrading command-and-control (C2) systems. This was followed on July 29 by a contract with the Ministry of Defense for "research and demonstration of AI functions needed for comprehensive analysis operations" — applying agent technology to the work of intelligence analysts at the Defense Intelligence Headquarters, moving the relationship from unit-level operations to national policy-level engagement.

The broader backdrop moved in step. Yomiuri reported on August 9 that the government had settled on a policy to introduce domestic AI into the Self-Defense Forces' command-and-control systems, with Sakana Fugu named as a leading candidate. The company maintains a separate defense-business policy document naming finance and defense/intelligence as priority sectors alongside each other. In 2025, at the U.S.-Japan Defense Innovation Challenge co-hosted by the Acquisition, Technology & Logistics Agency and the U.S. Defense Innovation Unit (DIU), the company won a commendation, becoming the only participant to reach the finals in both categories: pandemic prediction and AI-generated image detection. On May 29, it also formed an intelligence-analysis partnership with the general incorporated association DeepDive.

What It Still Has to Clear

The company's valuation is $2.7 billion, yet no revenue figures have been disclosed. The $30 million figure circulating in the market comes from third-party estimates and has never been confirmed by the company. Ha himself has said that generative AI has yet to produce a business model with proven profitability, and has raised the possibility of a broader valuation bubble across the industry. Given that Marlin, the company's first commercial product, only launched in June, there's still no external way to confirm whether revenue has caught up with valuation.

The competitive landscape is also far from simple. Google is both an investor and a competitor through Gemini in the same market. OpenAI, Anthropic, and Google DeepMind command orders of magnitude more compute and capital than Sakana AI. Retaining research talent means competing with them on compensation.

The company's focus on Japanese language and culture is both a moat and a ceiling. Every partner that has generated revenue so far has been a major Japanese corporation or government agency. Fugu is open to developers abroad through its OpenAI-compatible API and distribution via OpenRouter, but the European Economic Area remains blocked for now, and no figures on overseas revenue have been disclosed.

Questions about the technology's own reliability also remain. The AI CUDA Engineer's corrected performance claims, The AI Scientist's time-limit workaround, and the voluntary withdrawal of the workshop paper all raise the question of who verifies automated research systems — and how — before they're put into real-world use. For a company whose customers include banks and defense agencies, that question carries more weight than it would for ordinary software.

Editor's Take

There are two ways to read Sakana AI.

One is through its technical path. Every piece of this company's research starts from the same constraint: it doesn't have the GPU resources of America's big labs. Evolutionary model merging skips training entirely to save compute; NAMM cuts inference memory by up to 75%; ShinkaEvolve matches record performance using just 150 samples; and Fugu gets its performance by orchestrating other companies' models rather than building a new one. Resource scarcity shaped the research agenda itself. Given that most AI organizations in Korea face the same constraint, Sakana's way of framing problems is arguably more instructive than any single result it has produced.

The other lens is its business trajectory. Over three years, Sakana AI has climbed a ladder: research publication → corporate partnership → government/defense contract, with each stage built on the credibility established by the one before it. The three megabanks joining the Series A made the Mitsubishi UFJ contract possible; its track record in finance is why Citi made its first Japanese investment; and completing foundational research on command-and-control systems is why the Ministry of Defense contract followed four months later. That's a different sequence from opening a market with a single finished product.

The lack of disclosed revenue, the history of walking back performance claims once already, and heavy reliance on the Japanese market remain real risks. Given that the company is now expanding its research focus into physical AI and robotics, the things worth watching over the next year come down to two questions: whether Fugu and Marlin can generate revenue outside Japan, and whether the defense demonstrations turn into actual deployment contracts.