Read this if you buy AI for a Japan-based team or want one endpoint over many models. Sakana's pitch is orchestration plus Japanese-language strength, backed by a public research record; the catch is no consumer app and reliance on third-party model APIs.
Sakana AI is a Tokyo AI company with ex-Google and ex-Stability AI founders. Its main product, Fugu, is one API that orchestrates frontier models; Marlin handles long research jobs and Namazu covers Japanese-language text. Its latest funding round put it among Japan's most valuable AI startups.
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. Citi and Mitsubishi Electric made corporate strategic investments in 2026. · Valuation: $2.65B (Series B, Nov 2025)
About Sakana AI
Sakana AI was founded in Tokyo in July 2023 by David Ha, Llion Jones and Ren Ito. Ha was Research Director at Google Brain and later Head of Research at Stability AI. Jones co-authored the 2017 paper "Attention Is All You Need" at Google, which introduced the Transformer architecture. Ito was a Japanese diplomat, then an Executive Officer at Mercari from 2015, and COO of Stability AI in the UK. Ha is CEO, Jones is CTO and Ito is Chairman. The founders' thesis is that frontier performance can come from orchestrating many models and from nature-inspired learning, rather than from scale alone.
The 2026 lineup has three parts. Sakana Fugu (launched June 22, 2026) is a multi-agent orchestration system behind one OpenAI-compatible endpoint: a request either gets answered directly or is routed to a coordinated pool of frontier models. It ships as Fugu for low-latency everyday work and Fugu Ultra, a 1 million-token tier built for multi-step problems. Fugu Ultra is priced at $5 per million input tokens and $30 per million output tokens, and subscription plans run $20, $100 and $200 a month (Standard, Pro, Max). Sakana reports 73.7% for Fugu Ultra on SWE-Bench Pro, against 58.6% for GPT-5.5 and 54.2% for Gemini 3.1 Pro. Sakana Marlin (launched June 15, 2026) is a deep-research agent that works for up to eight hours on one brief and returns reports of up to 100 pages with slides; it builds on AB-MCTS and on The AI Scientist. Namazu is the family of Japanese-language models, including the compact TinySwallow-1.5B.
Funding totals about $379 million. The seed round was $30 million from Lux Capital and Khosla Ventures. The 2024 Series A raised roughly $214 million (about 30 billion yen) at a $1.5 billion valuation from Japanese banks and corporates including MUFG, SMBC, Mizuho, Itochu, KDDI and Nomura, plus Nvidia. The November 2025 Series B was $135 million (20 billion yen) at a $2.65 billion post-money valuation, with new investors Factorial Funds, Macquarie Capital, Santander through Mouro Capital and In-Q-Tel, and returning backers Khosla, NEA and Lux, alongside Salesforce Ventures, Google and Datadog. Citi made a strategic investment on February 24, 2026, its first in a Japanese company, and Mitsubishi Electric added a minority stake in March 2026, tied to its Serendie platform.
The research record is public: The AI Scientist appeared in Nature, AB-MCTS was a NeurIPS 2025 Spotlight, and the Darwin Godel Machine plus the TRINITY and Conductor multi-agent methods were presented at ICLR 2026. A Recursive Self-Improvement Lab opened in Tokyo in August 2026 to study systems that improve their own methods and code with less compute. Marlin ran a closed beta in April 2026 with about 300 finance and consulting professionals, and the company targets Japanese finance, consulting, manufacturing and government buyers.
For context, the nearest global names are OpenAI, Anthropic and Mistral AI; in Japan the closest Tokyo lab is Preferred Networks. The trade-offs are plain: there is no consumer product comparable to a chat assistant, and Fugu's quality depends on third-party frontier APIs, so a change in access to those models would change what Fugu can route to. If you are weighing it against another provider, Smart Match can narrow the choice by job.
Mission
Develop AI inspired by nature: evolution, collective intelligence, and sparsity, building sovereign AI technologies that capture Japanese language and culture.
Products
- Sakana Fugu (Orchestration model API): https://sakana.ai/fugu/
- Sakana Fugu Ultra (High-capacity orchestration model): https://sakana.ai/fugu/
- Sakana Marlin (Autonomous deep-research agent): https://sakana.ai/blog/
- Namazu LLMs (Japanese-language foundation models): https://sakana.ai
Sakana AI Models on Hokai
Links
Frequently Asked Questions
Where is Sakana AI based and who runs it?
Sakana AI is headquartered in Tokyo, and its founders are all veterans of Google or Stability AI. David Ha is CEO, Llion Jones is CTO and Ren Ito is Chairman. The team is roughly 150 to 200 people, all in Tokyo.
Who has invested in Sakana AI?
Backers include Lux Capital, Khosla Ventures, NEA, MUFG, Nvidia, Google, Salesforce Ventures, Macquarie Capital and In-Q-Tel. Citi and Mitsubishi Electric added strategic stakes in 2026. Japanese banks and corporates joined the earlier rounds.
Which Sakana AI product should I look at first?
Start with Fugu if you want one API that routes across frontier models, and with Marlin if the job is long-form research for a finance or consulting brief. Namazu fits Japanese-language text work on a small model. Each has its own page and pricing on sakana.ai.
What research is Sakana AI known for?
Its papers cover AI-driven scientific discovery, tree search that lets several models cooperate, and systems that rewrite their own code. The RSI Lab in Tokyo continues the self-improvement line.
How does Sakana AI differ from OpenAI or Anthropic?
OpenAI and Anthropic sell their own frontier models. Sakana's Fugu sits above models and routes among them, and its other pitch is Japanese-language and Japan-based supply. The cost is dependence on outside model APIs and no consumer assistant.
Top Alternatives
- OpenAI: Pick Sakana for routing across several models and Japan-based supply; pick OpenAI for its own frontier models and a consumer app.
- Anthropic: Pick Sakana when Japanese-language coverage or model-swapping matters; pick Anthropic for its own Claude models and enterprise safety posture.
- Mistral AI: Pick Sakana for multi-model orchestration and a Tokyo base; pick Mistral AI for open-weights models from a European lab.
Hokai guides that mention Sakana AI
- Best Autonomous AI Agents in 2026: How Much Control to Hand Over: The best AI agents of 2026 sorted by autonomy and oversight, with current prices for Claude Code, Jules, Hermes Agent, Fugu and Artisan, rechecked in November.