Sakana AI: $2.65B Japanese AI Startup (Founded 2023)

Sakana AI: Tokyo startup founded 2023, raised $379M at $2.65B valuation. Makes Fugu orchestration model, Marlin deep-research agent, and Namazu LLMs for Japan.

Sakana AI is a Tokyo-based startup founded in July 2023 by ex-Google and ex-Stability AI veterans. It has raised $379M at a $2.65 billion valuation and focuses on nature-inspired foundation models. Products include Fugu (an orchestration model routing tasks across frontier LLMs), Marlin (an 8-hour autonomous deep-research agent), and Namazu (Japanese-language LLMs).

Sakana AI is a Tokyo-based AI startup founded in July 2023 by David Ha (former Google Research Director), Llion Jones (co-author of 'Attention Is All You Need'), and Ren Ito (former COO of Stability AI). The company has raised $379M across three rounds at a $2.65 billion valuation. Its product lineup includes Fugu (multi-model orchestration API), Marlin (8-hour autonomous research agent), and Namazu (Japanese LLMs).

Founded: 2023 · HQ: Tokyo, Japan · Team: 150-200 · CEO: David Ha · Funding: $379M total raised across seed, Series A ($214M, 2024), and Series B ($135M, Nov 2025). Investors: Lux Capital, Khosla Ventures, MUFG, NEA, In-Q-Tel, Nvidia, Google, Salesforce Ventures. · Valuation: $2.65B (Series B, Nov 2025)

About Sakana AI

Sakana AI was founded in Tokyo, Japan in July 2023 by David Ha, Llion Jones, and Ren Ito. David Ha served as Research Director at Google Brain and later as Head of Research at Stability AI before co-founding Sakana. Llion Jones co-authored the 2017 paper "Attention Is All You Need" while at Google, the paper that introduced the Transformer architecture now central to nearly every major language model. Ren Ito spent years as a Japanese diplomat, then served as Executive Officer at Mercari from 2015, and was COO of Stability AI in the UK before joining the founding team. The three founded Sakana on the thesis that frontier AI performance can be achieved through model orchestration and nature-inspired learning rather than brute-force scale. Sakana AI's product lineup as of June 2026 includes three offerings. Sakana Fugu (launched June 22, 2026) is a multi-agent orchestration system exposed as a single OpenAI-compatible API: you send a request to one endpoint and Fugu decides whether to answer directly or route the query to a coordinated pool of frontier language models. It comes in two tiers: Fugu (low-latency, everyday tasks) and Fugu Ultra (1 million-token context, $5 per million input tokens and $30 per million output tokens, built for complex multi-step problems). On SWE-Bench Pro, Fugu Ultra scores 73.7%, ahead of GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%. Sakana Marlin, launched June 15, 2026, is an autonomous deep-research agent that works for up to eight hours on a single strategic brief and delivers reports of up to 100 pages with slides. Marlin is built on AB-MCTS (Adaptive Branching Monte Carlo Tree Search), a NeurIPS 2025 Spotlight result, and on The AI Scientist framework published in Nature. Pricing is 100 credits per run or monthly plans starting at roughly $1,000. The third product line, Namazu, is Sakana's family of Japanese-language LLMs including TinySwallow-1.5B, a compact model for Japanese text tasks. Sakana AI raised $30 million in its seed round from Lux Capital and Khosla Ventures. In 2024 the company raised approximately $214 million (roughly 30 billion yen) in its Series A at a $1.5 billion valuation, with participants including Mitsubishi UFJ Financial Group, SMBC, Mizuho, Itochu, KDDI, Nomura, and Nvidia. In November 2025, Sakana closed a $135 million (20 billion yen) Series B round at a $2.65 billion post-money valuation; new investors included Factorial Funds, Macquarie Capital, Santander Group via Mouro Capital, and In-Q-Tel (the CIA's venture arm), joining returning investors Khosla Ventures, NEA, Lux Capital, plus Salesforce Ventures, Google, and Datadog. Mitsubishi Electric made an additional corporate minority investment in March 2026. Total funding across all rounds is approximately $379 million. Sakana AI earns revenue from Fugu subscription plans at $20/month (Standard), $100/month (Pro), and $200/month (Max), as well as pay-as-you-go API pricing. The company's primary enterprise targets are Japan's finance, consulting, manufacturing, and government sectors. Marlin's closed beta in April 2026 ran with approximately 300 professionals from finance and consulting, validating demand before the general release. Public ARR estimates suggest roughly $30 million as of 2025, though Sakana has not officially confirmed a revenue figure. David Ha serves as CEO, Llion Jones as CTO, and Ren Ito as Chairman. The team has grown to approximately 150-200 employees as of mid-2026, all based in Tokyo. Sakana published The AI Scientist in Nature and presented AB-MCTS as a NeurIPS 2025 Spotlight paper. The company also introduced the Darwin Gödel Machine (a self-rewriting AI system) and the TRINITY and Conductor frameworks (multi-agent coordination methods grounding the Fugu architecture), both presented at ICLR 2026. Sakana AI's stated mission is to develop AI based on nature-inspired intelligence: evolution, collective intelligence, and sparsity. The research agenda distinguishes Sakana from labs focused purely on scaling: rather than training ever-larger single models, Sakana studies how populations of smaller models can collaborate to solve problems that exceed any individual model's capability. This approach also addresses a practical concern: Japan-based enterprises face uncertainty about long-term access to potentially export-controlled US frontier models, and Fugu's swappable model pool is designed so that any specific model can be removed or replaced without changing the API interface. Sakana's primary competitors in the orchestration and Japanese LLM spaces include OpenAI (dominant globally but without Japan-specific models), Anthropic (strong on safety and enterprise, no Japan focus), and Preferred Networks (a Tokyo-based AI lab building Japanese industrial AI). Sakana beats all three on Japan-language model quality and on sovereign AI positioning for Japanese enterprises. Where Sakana is weakest: it has no consumer product comparable to ChatGPT or Claude.ai, and its orchestration layer depends on third-party frontier model APIs, so changes to OpenAI or Anthropic API access could affect Fugu's model pool.

Mission

Develop AI inspired by nature: evolution, collective intelligence, and sparsity, building sovereign AI technologies that capture Japanese language and culture.

Products

Links

Website · GitHub · Twitter · LinkedIn · Blog · Docs

Frequently Asked Questions

What is Sakana AI and what do they build?

Sakana AI is a Tokyo-based AI startup founded in July 2023 that builds foundation models and AI products for Japanese and global enterprise markets. The company was created by David Ha (former Head of Research at Stability AI and Research Director at Google Brain), Llion Jones (CTO and co-author of the original Transformer paper), and Ren Ito (Chairman and former COO of Stability AI). Its flagship product as of June 2026 is Sakana Fugu, an orchestration model that routes user queries across a pool of frontier language models through a single OpenAI-compatible API, removing the need for users to pick or manage individual models. Fugu Ultra, the performance tier, achieved 73.7% on SWE-Bench Pro, ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). Sakana also offers Marlin, an autonomous deep-research agent that produces 100-page reports in up to 8 hours, and Namazu, a family of Japanese-language LLMs. All products can be accessed via sakana.ai or the developer console at console.sakana.ai.

Who founded Sakana AI and who is the CEO?

Sakana AI was founded in July 2023 by three co-founders from prior leadership roles at major AI organizations. David Ha is CEO; before Sakana, he served as Head of Research at Stability AI and as a Research Scientist and Research Director at Google Brain, where he published influential work on neural evolution and world models. Llion Jones is CTO and is one of the eight original co-authors of the 2017 paper 'Attention Is All You Need,' which introduced the Transformer architecture used in virtually all major LLMs today. Ren Ito is Chairman; his career includes roles as a Japanese diplomat, Executive Officer for global business at Mercari from 2015, and COO of Stability AI in the UK from 2022. All three left their prior organizations with the shared belief that nature-inspired coordination methods could achieve frontier performance without requiring the largest possible single model. Sakana is incorporated as a Japanese entity and its team works entirely out of Tokyo. No major leadership changes have been reported since founding.

How much funding has Sakana AI raised?

Sakana AI has raised approximately $379 million in total across three main funding rounds. The seed round raised $30 million from Lux Capital and Khosla Ventures shortly after founding in 2023. The Series A raised approximately $214 million (roughly 30 billion yen) in 2024 at a $1.5 billion valuation, with investors including Mitsubishi UFJ Financial Group, SMBC, Mizuho, Itochu, KDDI, Nomura, and Nvidia; this round made Sakana a unicorn in its first full year of operation. The Series B raised $135 million (roughly 20 billion yen) in November 2025 at a $2.65 billion post-money valuation, adding new investors Factorial Funds, Macquarie Capital, Santander Group via Mouro Capital, and In-Q-Tel (the CIA's venture capital arm), alongside returning investors Khosla Ventures, Lux Capital, NEA, Salesforce Ventures, Google, and Datadog. Mitsubishi Electric made an additional corporate minority investment in March 2026. No IPO timeline has been announced, though the company has signaled interest in strategic partnerships and M&A for global expansion.

What products does Sakana AI make?

Sakana AI offers three main product lines as of June 2026. Sakana Fugu, launched June 22, 2026, is a multi-agent orchestration system packaged as a single OpenAI-compatible model API; it internally routes queries to the most suitable combination of frontier language models and returns one synthesized answer, with the user never managing any underlying model. Fugu comes in two tiers: Fugu (lower latency, everyday tasks) and Fugu Ultra (1 million-token context window, priced at $5 per million input tokens and $30 per million output tokens). Subscription plans are available at $20/month (Standard), $100/month (Pro), and $200/month (Max). Sakana Marlin, launched June 15, 2026, is a B2B autonomous research agent that accepts a strategic brief and autonomously researches and synthesizes it for up to eight hours, producing reports of up to 100 pages with slides; it is priced at 100 credits per run or approximately $1,000 per month for a subscription. The third line, Namazu, is a family of Japanese-language LLMs including TinySwallow-1.5B, a compact model for Japanese text tasks used by enterprise customers in finance and manufacturing.

Where is Sakana AI headquartered and how big is the team?

Sakana AI is headquartered in Tokyo, Japan and operates from a single office location in Japan. The company has no confirmed overseas offices as of June 2026. Team headcount estimates from third-party trackers in mid-2026 range from approximately 150 to 200 employees, with Tracxn reporting 199 employees as of May 31, 2026. The Japan Times reported in September 2025 that the team numbered more than 100 including contractors. Team growth has been consistent with the fundraising pace: a small founding team in mid-2023 scaled to over 100 by late 2024 and continued growing through the 2025 Series B. The company hires researchers and engineers with backgrounds in machine learning, Japanese NLP, and distributed systems. Sakana has advertised roles in applied research engineering as of 2026.

What is Sakana AI's mission and research focus?

Sakana AI's mission is to develop AI inspired by nature: specifically, principles of evolution, collective intelligence, and sparsity. The company's core research thesis is that a coordinated population of smaller, specialized models can match or exceed a single massive model on complex tasks at lower compute cost. Key published research includes The AI Scientist (a system that automates the scientific discovery process, published in Nature), the Darwin Gödel Machine (a self-rewriting AI that improves its own code), AB-MCTS (Adaptive Branching Monte Carlo Tree Search, a NeurIPS 2025 Spotlight paper), and the TRINITY and Conductor frameworks for multi-agent coordination (both presented at ICLR 2026). The Namazu Japanese LLMs address the company's focus on culturally grounded AI for Japanese enterprise. Sakana's approach also has a commercial rationale: by not relying on any single frontier model, Fugu's architecture protects Japanese customers from potential export-control restrictions on US-based models. The company publishes its research openly on arXiv and in top conferences including NeurIPS and ICLR.

Is Sakana AI compliant with SOC 2, GDPR, or HIPAA?

As of June 2026, Sakana AI has not publicly confirmed any compliance certifications including SOC 2, ISO 27001, GDPR data processing agreements, or HIPAA eligibility. The company does not appear to have a published trust center or security documentation page as of this writing. API access is managed through the Sakana console at console.sakana.ai, and API documentation is available at the Sakana developer portal. Enterprises requiring formal compliance certification should contact Sakana AI directly to request current security documentation. The company stores data in Japan, which may be relevant for customers with Japanese data sovereignty requirements, but no formal regional compliance certifications have been disclosed. Given that Sakana's target markets include Japan's financial sector and government, it is likely that compliance certifications are being pursued, but none have been announced publicly. This field should be re-verified when Sakana publishes a trust center or security page.

Who are Sakana AI's main competitors?

Sakana AI competes primarily in three areas: multi-model orchestration APIs, Japanese-language LLMs, and autonomous AI agents. On orchestration, the closest direct competitor is OpenAI (GPT-5.5, 58.6% on SWE-Bench Pro vs Fugu Ultra's 73.7%), though OpenAI does not offer an orchestration-layer product; Anthropic's Claude Opus 4.8 (69.2% on SWE-Bench Pro) is another benchmark reference point. On Japanese LLMs, Preferred Networks is the most prominent Tokyo-based competitor building Japanese industrial AI, though it targets manufacturing robotics more than enterprise text tasks. On deep-research agents, Sakana Marlin competes with Perplexity's Deep Research, Google's AI Co-Scientist, and similar long-horizon research products. Sakana's clearest advantage over all three competitor categories is its Japan-sovereign positioning: Fugu is explicitly designed to operate without dependency on US export-controlled models. Its main weakness versus all of these competitors is the absence of a consumer-facing product and a smaller overall developer community than OpenAI or Anthropic.