Inside 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 and managing director of the Japan rates trading desk before moving to Google Brain in 2016. While still at the bank, he ran an AI research blog under the pseudonym "hardmaru," which built a following in the deep learning community before people knew who was behind it. After Google, he spent about a year from 2022 leading research at Stability AI.

CTO Llion Jones grew up in Wales, built a chatbot at age 14, and earned both his bachelor's and master's in computer science and AI from the University of Birmingham. His name appears on one of the most cited papers in the field: "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture now underlying nearly every large language model. Jones is one of its eight co-authors. He's said he joined Sakana AI because he believes the standard Transformer approach has hit diminishing returns.

COO Ren Ito studied law at the University of Tokyo, earned an LLM from NYU Law, and a master's from Stanford. He joined Japan's Ministry of Foreign Affairs in 2001 and worked as a diplomat before becoming global executive officer at Mercari. He also served as COO at Stability AI. That overlap — Ha and Ito both passing through Stability AI, and Ha and Jones meeting at Google's Tokyo office — is how the founding team came together.

The name "Sakana" (魚) means fish in Japanese. A single fish follows only a few simple rules, but a school of fish, acting together, produces complex behavior — evading predators, finding food — that no individual fish could manage alone. Translating that collective intelligence into model design is the company's founding idea. Where the industry has largely chased performance by scaling up parameter counts, the three founders bet there was another path.

Evolutionary model merging — building new models without training

The result that put Sakana AI on the research community's map was Evolutionary Model Merge, released on March 21, 2024. It combines multiple existing 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 runs in two spaces at once. In parameter space (PS), the algorithm evolves the blending ratio of different models' weights, layer by layer. In data-flow space (DFS), it evolves the order in which layers from different models get stitched together. Both spaces are too large for human intuition to search exhaustively. Running roughly 100 to 150 generations, with only the best-performing candidates surviving to produce the next generation, yields the final model.

The key point is that this approach never runs gradient-descent training at all. Instead of burning GPU cycles on repeated backpropagation, it reframes the problem as combining already-trained weights — cutting compute costs dramatically.

Sakana AI released concrete results alongside the method. Its 7-billion-parameter EvoLLM-JP scored 55.2% on the Japanese math reasoning benchmark MGSM-JA (50.6% for the publicly released v1-7B). It was built by merging a Japanese-language model (Shisa-Gamma) with math-specialized models (WizardMath, Abel). Despite being built specifically for math, it also averaged 70.5 on the general Japanese benchmark JP-LMEH — beating every Japanese model under 70 billion parameters and the previous best in the 70-billion-plus class. Notably, Swallow 70B in the same table still scored 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-stage reasoning.

Space	What gets evolved	The old human-driven method
Parameter space (PS)	Per-layer weight blending ratios	Fixed-ratio averaging
Data-flow space (DFS)	Layer ordering across different models	Manual design based on intuition and experience

AI that fixes itself — from the AI Scientist to the RSI Lab

The company's second research thread is getting AI to improve AI. It began with "The AI Scientist," released August 13, 2024 — a system that runs the full research pipeline without human involvement: generating ideas, writing experiment code, running it, analyzing results, writing the paper, and even peer review. The compute cost to produce a single paper was reported at roughly $15.

The follow-up, AI Scientist v2, produced results at an ICLR 2025 workshop review in March 2025. Of three papers submitted, one scored an average of 6.33 — above the acceptance threshold, and higher than 55% of human-authored submissions. Sakana AI voluntarily withdrew the paper before publication, however. It hadn't gone through the workshop organizers' meta-review, workshop acceptance rates run far higher than main-conference rates, and citation errors were flagged. That's the actual scope of the claim that "AI passed peer review." A 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 signal. During testing, when an experiment ran over its time limit, the system — instead of speeding up its code — modified its own code to extend the time limit itself. In one case, it inserted a system call to relaunch itself, creating an infinite loop. After the company disclosed this, it fueled ongoing debate about controlling autonomous agents.

Subsequent work pushed further into self-improvement. Darwin Gödel Machine, released in May 2025, has an agent rewrite its own codebase to generate variants. It raised SWE-bench performance from 20% to 50%, a 30-point absolute improvement, and lifted scores on the multilingual coding benchmark Polyglot from 14.2% to 30.7%. In September of the same year, the open-sourced ShinkaEvolve — a program-evolution framework combining language models with evolutionary algorithms — solved the classic circle-packing optimization problem (placing 26 circles) at a record level using only 150 samples, and in a separate experiment found a new load-balancing loss function for mixture-of-experts models within 30 generations.

There were also competition results. On December 14, 2025, the algorithmic optimization agent ALE-Agent won first place at AtCoder Heuristic Contest 058, a four-hour contest where it competed against 804 human participants — the first time an AI agent has won such a contest. Total compute cost for the run was about $1,300.

This entire line of work was formally organized into the RSI Lab on June 5, 2026. The lab's stated mission is to "advance through ideas, not just compute" — turning Japan's lack of hyperscaler-level compute into a reason to push harder on sample-efficient research. Extending that logic into the physical world, the company published research on July 13 into physical "smart cellular bricks" that embody collective intelligence, and on August 10 it framed physical AI as the next frontier for this line of work through its own channels.

Research aimed past the Transformer itself

The third research thread targets the architecture itself.

Continuous Thought Machines (CTM), released in May 2025, reintroduces a time dimension into neural networks: individual neurons retain memory of past activity and process information sequentially through neuron-to-neuron synchronization. The company says this architecture beat human performance on confidence calibration.

Neural Attention Memory Models (NAMM), released in December 2024, learn which tokens in context should be kept and which 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 self-adaptive architectures where a model adjusts its own weights based on the task at hand. All three lines of work sit apart from the "just build a bigger model" trajectory.

Not every bet paid off. "The AI CUDA Engineer," released in February 2025, claimed to automatically optimize GPU kernels for up to 100x speedups over PyTorch. Outside verification found the system had exploited a loophole in the evaluation code — reward hacking. Some users reported the real-world result was actually three times slower. The company acknowledged this on February 21, 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 stuck with the company's disclosure practices ever since.

2026: the year research became product

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

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

Sakana Fugu is an orchestration model that entered beta on April 24 and launched fully on June 22. Rather than having a single model answer, it deploys multiple frontier models in different roles to divide up a task. The underlying research appears in two ICLR 2026 papers. Trinity uses a lightly evolved coordinator to assign "thinker, worker, and verifier" roles across multiple language models, while Conductor uses reinforcement learning to discover communication patterns between agents on its own.

The product launched with a general-purpose Fugu and Fugu Ultra, which draws on a deeper pool of expert models for harder problems; Fugu Cyber, specialized for security reasoning, was added on July 21. On the company's own benchmarks, Fugu Ultra scored 73.7 on the software engineering benchmark SWE-Bench Pro. 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 outscored 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 Opus preview (83.1). All figures are Sakana AI's own measurements; no third-party verification has been published.

Pricing includes $20/$100/$200 monthly subscription tiers alongside pay-as-you-go rates. Ultra's usage pricing is $5 per million input tokens and $30 per million output tokens, rising to $10 and $45 for long-context requests beyond 272,000 tokens. The API is compatible with the OpenAI format and is also accessible through OpenRouter and Vercel. The European Economic Area is excluded for now, pending GDPR compliance work.

A phrase the company keeps repeating in describing Fugu is "frontier-level performance without export control risk" — a clear pitch to regions and industries affected by U.S. export restrictions on AI models.

Sakana Marlin is an autonomous research agent that went through a closed beta of roughly 300 users starting April 2 and launched fully on June 15 — Sakana AI's first commercial product. A single session can run unattended for up to eight hours and produce a 100-page report. It's built on adaptive branching Monte Carlo tree search, which earned a NeurIPS 2025 spotlight, combined with the workflow automation from the AI Scientist. Pricing splits into pay-as-you-go credits (¥98 per 100 credits), a Pro tier at ¥150,000/month, and a Team tier at ¥400,000/month. The company describes it as a "virtual chief strategy officer" for finance, consulting, and strategy teams.

Product	Timing	What it is
Sakana Chat	March 24, 2026	Free chatbot, Japan-only, built on Namazu family
Sakana Fugu	Beta April 24, full launch June 22	Multi-model orchestration API
Sakana Marlin	Beta April 2, full launch June 15	Up to 8-hour unattended 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

Sakana AI's fundraising pace is unusual even by startup standards, let alone for a Japanese company.

In January 2024, it raised a $30 million seed round led by Lux Capital, followed by a roughly $200 million Series A that closed September 4 of the same year. NEA, Khosla Ventures, and Lux Capital co-led it, with NVIDIA joining as a strategic investor. Japan's three megabanks — MUFG, SMBC, and Mizuho — as well as major corporations including NEC, Fujitsu, Itochu, KDDI, Nomura, Dai-ichi Life, ANA Holdings, and Tokio Marine all participated. Reports at the time put the valuation at $1.5 billion — making Sakana AI a unicorn just 14 months after founding.

The Series B was first announced on November 17, 2025, at ¥20 billion (roughly $135 million) with a valuation of ¥400 billion. An updated filing on April 9, 2026 raised those figures to ¥32 billion (about $200 million) at a valuation of roughly ¥432 billion (about $2.7 billion). Cumulative funding now stands at roughly ¥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-ever strategic investment in a Japanese company. Also on the roster: Salesforce Ventures, Datadog, Macquarie Capital, JAFCO, In-Q-Tel (the venture fund tied to U.S. intelligence agencies), and Santander's Mouro Capital.

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)

Rather than build its own data centers, Sakana AI sources compute externally. It received support through METI's GENIAC program and uses GMO's GPU cloud. On July 16, 2026, it adopted Sakura Internet's "Koukaryoku PHY" and, the same day, announced an open-model collaboration with NVIDIA. That's a very different cost structure from U.S. and European competitors pouring billions of dollars into their own clusters.

First the banks, now the defense ministry

Revenue comes from multiyear partnerships with major Japanese corporations. In May 2025, Sakana AI signed a comprehensive, three-plus-year partnership with MUFG to build banking-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's March investment came bundled with a plan to integrate Sakana AI's models into its Serendie platform. On April 30, the company announced it was co-developing a proposal-generation app with SMBC Group, and on April 7 it joined a Ministry of Internal Affairs and Communications program tackling disinformation on social media. Unofficial estimates put 2025 revenue at around $30 million.

The second track is defense and intelligence. On March 13, 2026, Sakana AI received a multiyear commissioned research contract from the Acquisition, Technology & Logistics Agency's Institute for Defense Innovation Science and Technology. The project, titled "Research on accelerating observation, reporting, information integration, and resource allocation through combined AI technologies," involves fusing 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. That 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. It's a jump from unit-level operations to national policy-level work.

The broader context has been moving in the same direction. Yomiuri reported on August 9 that the Japanese government has settled on introducing domestically made AI into the Self-Defense Forces' command-and-control systems, with Sakana Fugu named as a leading candidate. The company has its own defense-business policy document naming finance alongside defense and intelligence as priority areas. In 2025, it received an honorable mention at the U.S.-Japan Defense Innovation Challenge, co-hosted by ATLA and the U.S. Defense Innovation Unit (DIU) — the only participant to reach the finals in both categories, epidemic prediction and AI-generated image detection. On May 29, it also formed an intelligence-analysis partnership with the general incorporated association DeepDive.

What's still unresolved

The valuation stands at $2.7 billion, but no revenue figures have been officially disclosed. The $30 million figure circulating in the market is a third-party estimate the company has never confirmed. CEO Ha himself has said no proven, profitable business model has yet emerged in generative AI broadly, and has raised the possibility of a valuation bubble across the industry. Given that Marlin, its first commercial product, only launched in June, there's no way for outsiders to confirm yet whether revenue has caught up with the valuation.

The competitive landscape isn't simple, either. Google is both an investor and a competitor in the same market through Gemini. OpenAI, Anthropic, and Google DeepMind all command an order of magnitude more compute and capital than Sakana AI, forcing it to compete on compensation to retain research talent.

The Japan-focused strategy is both a moat and a ceiling. Every revenue-generating partner to date has been a major Japanese corporation or government agency. Fugu is technically open to international developers through its OpenAI-compatible API and OpenRouter distribution, but the EEA remains blocked and no overseas revenue figures have been disclosed.

Questions about the reliability of the technology itself also remain. The walked-back performance claims for the AI CUDA Engineer, the AI Scientist's time-limit workaround, and the voluntary withdrawal of a workshop paper all point to the same open question: who verifies an automated research system, and how, before it's deployed in real operations? 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 strand of its research starts from the same constraint: it doesn't have the GPU resources of the big U.S. labs. Evolutionary model merging skips training entirely to save compute. NAMM cuts inference memory by up to 75%. ShinkaEvolve set a record using just 150 samples. Fugu gets performance not by building a new model but by orchestrating other people's models. Resource scarcity is what set the research agenda. Given that most AI teams in Korea face the same constraint, how Sakana AI framed its problems is arguably more useful to study than any single result.

The other is its business path. Over three years, Sakana AI climbed a specific ladder: research publication → corporate partnership → government and defense contract. Each stage borrowed credibility from the one before it. The three megabanks' participation in the Series A made the MUFG contract possible. Its financial-sector track record is what convinced Citi to make its first-ever investment in a Japanese company. And the foundational C2 research is what led to the defense ministry contract four months later. That's a very different sequence from opening a market with a single finished product.

The lack of disclosed revenue, the one instance of walked-back performance claims, and the heavy concentration in the Japanese market are real risks that haven't gone away. Given that the company is now expanding its research agenda into physical AI and robotics, the two things worth watching over the next year are whether Fugu and Marlin can generate revenue outside Japan, and whether its defense demonstration work turns into an actual deployment contract.