Home › Blog › Sakana AI: A Primer on Tokyo's Nature-Inspired Frontier Lab
Company Primer

Sakana AI: A Primer on Tokyo's Nature-Inspired Frontier Lab

Founded in 2023 by a Transformer co-author and a former Google Brain Japan lead, Sakana AI bets that evolving and combining models will beat simply scaling one. This primer covers where it came from, who is behind it, what it has shipped, which papers matter, and which claims are still only the company's own.

FL
FrontierAGI Team
⚠️ How this was researched Everything here comes from web searches. Sakana's own pages and the arXiv papers could not be opened directly while writing, so figures that trace to the company, a paper abstract or a news report are labelled as such, and the confidence badges in the timeline say how well each item is documented. Where sources disagree (the Series B amount is the clearest case) both numbers are shown instead of one being picked. Anything we could not confirm is called out in the section on conflicting numbers.

Sakana AI at a glance

Sakana AI is a Tokyo-based AI research company founded in 2023 by David Ha, Llion Jones and Ren Ito. Unlike the labs that compete on the largest possible training run, it builds around ideas borrowed from nature: evolutionary search, collective behaviour and systems that improve themselves. In 2026 it began shipping commercial products and formalised a lab dedicated to recursive self-improvement.

2023Founded in Tokyo by David Ha, Llion Jones and Ren Ito
~$2.65BValuation after the Nov 2025 Series B, per CEO David Ha as reported by TechCrunch
¥32B vs ¥20BSeries B size: Sakana's own announcement says ¥32B (about $200M); most press says ¥20B (about $135M)
NatureThe AI Scientist was published in Nature in March 2026; model merging in Nature Machine Intelligence in Jan 2025

Foundation: why it exists

Sakana's founding pitch was a contrarian one. The dominant strategy in 2023 was to build one very large model and keep scaling it. David Ha and Llion Jones, both of whom had spent years at Google, argued for something closer to biology: many smaller models, evolved and combined, from which capable behaviour emerges. The company's announcement said it aimed to create a new kind of foundation model based on nature-inspired intelligence.

The name carries the idea. Sakana (さかな, 魚) is Japanese for fish, and the founders describe a school of fish: individual fish are simple, but together they move as one coherent whole. Llion Jones has given the same explanation in interviews, saying an individual fish might not be very smart, but a school can show emergent behaviour. The logo shows a school forming one shape with a single red fish swimming the other way; the company says the red fish stands for not simply doing what everyone else does.

Two other strands shaped the company. The first is the founders' pedigree. Jones is a co-author of the original "Attention Is All You Need" paper that introduced the Transformer. Ha led the Japan research team at Google Brain and is known for work on world models, evolution strategies and self-organising systems. The second is Japan. Sakana is positioned as a Japanese AI champion, with models and products tuned to Japanese language and culture and customers drawn from Japanese finance and industry.

💡 The core bet Instead of paying for ever-larger training runs, combine and evolve what already exists, and let AI systems do more of the research themselves. Every major Sakana project, from model merging to the AI Scientist to the RSI Lab, is a variation on that idea.

History: from a fish logo to a unicorn

Click an event for what happened, how sure we are, and where it comes from.

Sakana AI timeline, 2023 to 2026

Scroll sideways on a phone. Confidence badges say how firmly each date or claim is documented; "company-reported" means the claim comes from Sakana's own announcement.

Key people

Three co-founders run the company, and a wider circle of researchers and academic collaborators appears on its major papers. Click a card for the detail.

Who is behind Sakana

Titles follow Sakana's own company page where we could confirm them. Sources differ on Ren Ito's title, so both are given.

Main purpose and goal

Sakana's stated purpose has moved in three steps, and the order matters for understanding what it is today.

  1. Nature-inspired foundation models (2023 to 2024). Build models by evolving and merging, with collective intelligence as the guiding metaphor.
  2. Automating research (2024 to 2025). If evolution can search model space, it can search ideas too. The AI Scientist and the ALE-Agent programming agent aim at AI that does research and algorithm design itself.
  3. Recursive self-improvement (2026). Systems that improve not only their outputs but the process that produces future improvements. In June 2026 Sakana set up a dedicated RSI Lab around that goal.

The commercial side runs in parallel. Sakana says it plans to use its Series B money for research and engineering, plus sales and distribution in Japan, and to extend beyond finance into industrial, manufacturing and government customers. Its stated edge in the RSI Lab is sample efficiency: it wants the most efficient self-improvement engine, not the most compute-hungry one. That thesis is the company's own and has not been independently tested.

How the pieces fit together

Sakana's projects look scattered but form a lineage. Each rung reuses the one before it, and the RSI Lab pulls them under one roof. Click a node.

Four tracks leading to the RSI Lab

Horizontal position is roughly the year of release. Lines show how one project fed the next; they are our reading of the public record, not a company diagram.

Key deliverables

Three kinds of output: research systems, models, and, since 2026, commercial products.

What Sakana has shipped

Product dates marked "unofficial" come from third-party profiles we could not match against Sakana's own pages.

Key papers

Sakana publishes heavily and often with academic partners, which gives its claims more outside checking than a typical company blog post. The table is filterable by theme.

Seven papers to know

Venue and result lines are taken from abstracts, paper pages and news coverage; we did not read the full texts.

PaperVenue / dateWhat it showedCaveat

Funding and valuation

Sakana has raised three announced rounds and a strategic investment from Google. Toggle the Series B figure to see the two reported amounts.

Funding rounds (US$ millions, as announced or reported)

The Series A bar is the "over $100 million" announced in September 2024; the company later said it expanded, and Japanese coverage puts the total at about ¥30B.

Seed: VentureBeat. Series A: Maginative. Series B: Sakana and TechCrunch.
Seed (Jan 2024)
$30M, led by Lux Capital with Khosla Ventures; VentureBeat also lists Sony, NTT and KDDI.
Series A (Sep 2024)
Over $100M announced, led by NEA, Khosla Ventures and Lux Capital with NVIDIA taking part; later expanded with Japanese banks and corporates, about ¥30B in total per Japanese coverage. A $1.5B valuation appears in aggregators only.
Series B (17 Nov 2025)
Investors named in coverage include MUFG, Khosla Ventures, Macquarie Capital, NEA, Lux Capital and In-Q-Tel; a valuation of about $2.65B was given by Ha. One aggregator says MUFG led; we found no second source.
Google (Jan 2026)
Strategic partnership plus an equity investment, amount undisclosed. Sakana will build on Gemini and Gemma and target Japanese finance and government.
Total raised
Aggregators range from about $228M to $412M depending on which rounds they count. Do not quote a single figure.
Secondary market
One tracker, citing Hiive, showed a last price implying roughly $1.8B in September 2026, below the Series B mark. Treat it as a sentiment indicator, not a valuation.

The RSI Lab, and where Sakana sits on our map

On about 5 to 6 June 2026 Sakana announced the Sakana AI RSI Lab, a dedicated research group in Tokyo working on recursive self-improvement. It is not a separate company or funding round. The announcement frames the Darwin Gödel Machine, The AI Scientist, ShinkaEvolve and ALE-Agent as its foundations, and says the aim is open-ended systems that collectively improve themselves under constraints. On 24 September 2026 Sakana said Jürgen Schmidhuber, a long-time theorist of self-improving systems, would join as Chief Scientific Advisor and help guide the lab; coverage says he keeps his existing positions.

For readers of this site: the Darwin Gödel Machine is in the catalog behind our RSI Evolution Map at autonomy level L5 (it improves the mechanism that produces future improvements) in the harness family, because it rewrites its own coding-agent code and keeps an archive of variants. For the wider picture, see the RSI primer and AI automating AI research. Sakana's releases also appear on the Frontier Lab Updates page.

A critical view: what is shown and what is claimed

Sakana's headlines are strong. The evidence behind them varies a lot, so here is a ladder from peer-reviewed to purely company-reported.

How well is each headline supported?

Our own rough rating, 0 to 4: 4 = peer-reviewed venue; 3 = accepted or externally checkable; 2 = preprint or public logs from the company; 1 = company statement only; 0 = not disclosed.

🔍 Points to keep in mind
  • The AI Scientist's workshop paper. One of three AI-written papers was accepted at a workshop of ICLR 2025 (scores 6, 7, 6) and then withdrawn by the authors, as pre-agreed. Workshops accept far more papers than main conferences, and Sakana itself has said the paper would not meet its bar for a top venue. The Nature paper's authors also note the system can produce naive or underdeveloped ideas.
  • Darwin Gödel Machine numbers. SWE-bench 20.0% to 50.0% and Polyglot 14.2% to 30.7% are from the paper's abstract. Secondary write-ups describe reward-hacking behaviour in the experiments; the accounts differ and we did not confirm them in the paper.
  • ALE-Agent's contest win. The first-place result in AtCoder Heuristic Contest 058 (14 Dec 2025, 804 human participants, about $1,300 of compute) is Sakana's own account, backed by public contest standings and logs.
  • Business. Revenue and customer counts have not been disclosed. Partnerships with MUFG and Daiwa Securities are announced, and the products are new. One third-party site claims a defence contract; we could not corroborate it and have not used it.
  • RSI Lab. A stated mission and a hiring drive, not yet results. No independent evaluation of the lab exists.

Reading the numbers: where sources disagree

ItemWhat you may seeWhat the sources say
Series B size$135M / ¥20BSakana's announcement: ¥32B (about $200M). TechCrunch, Investing.com and others: ¥20B (about $135M). We found nothing explaining the gap, so we show both.
Series A$214M, $70M, $100M+The company announcement and Japanese reporting give over $100M, later about ¥30B in total. The $214M and $70M figures come from low-quality aggregators.
Total raised$228M, $302M, $379M, $412MEach tracker counts different rounds. No consistent total is documented.
Series B valuation$2.6B, $2.65B, $2.7B$2.65B post-money from Ha via TechCrunch is the best-sourced figure.
Fugu launchApril or June 2026One source says beta in April and general availability in June; another says June. Unresolved.
Namazu APILaunched 3 Aug 2026Only an unofficial third-party profile says so; we could not find Sakana's own announcement.
AI Scientist costAbout $15 per paperThat figure is for version 1. Version 2 is reported at roughly $20 to $25 in secondary blogs, and sources disagree.
Ito's titleCOO or ChairmanSakana's page lists Chairman; a Salesforce Ventures listing says COO.

🔗 Reference links

🧭 Closing

🎯 The bottom line Sakana AI is one of the few frontier-adjacent labs with a clear, consistent thesis, peer-reviewed work in Nature and Nature Machine Intelligence, and a strategy built on efficiency over scale. What is proven so far is research: model merging, an AI-written workshop paper, a self-modifying coding agent. What is still unproven is whether those ideas turn into a durable business or into self-improving systems that beat the labs spending far more on compute. Watch the RSI Lab's first independent results and the first disclosed revenue figures.