Fugu AI - Japanese Sakana AI Model (FUGU)

Sakana AI: Japan's nature-inspired AI lab

Sakana AI is a Tokyo-based artificial intelligence research lab founded in 2023 by David Ha and Llion Jones — both former Google researchers, with Jones known as a co-author of the landmark "Attention Is All You Need" paper that introduced the transformer architecture. The name means "fish" in Japanese, a nod to the lab's founding idea: that AI can learn from nature's own design principles.

A different philosophy from the giants

While the biggest labs race to train ever-larger foundation models, Sakana AI has pursued a distinctive research direction: evolution-inspired AI. Their work explores how techniques borrowed from biology — evolution, collective behavior, adaptation — can produce capable AI systems more efficiently than brute-force scaling. That includes research into automatically combining and evolving open models into stronger ones, and building systems where many smaller models work together.

The AI Scientist

Sakana's most talked-about project is the AI Scientist, a framework for automating the scientific research process itself — generating ideas, running experiments, and writing up results. First presented in 2024, it sparked wide debate about the future of automated discovery: exciting for accelerating research, controversial for what it implies about peer review and scientific labor. A follow-up version extended the concept further.

What "nature-inspired" means in practice

The slogan becomes concrete in their research program. Instead of training one giant model from scratch — the most expensive path — Sakana's researchers have explored evolutionary model merging: treating existing open models as a gene pool, combining them, and selecting the offspring that perform best. It's a genuinely different optimization philosophy: search and recombination over gradient descent alone. Whether it scales to the frontier is an open question, but it's produced real, published results and a distinctive identity in a field where most labs' methods look increasingly alike.

The open-model connection

A notable feature of Sakana's approach is building on the open model ecosystem — community and openly released models as raw material for research. This positions them differently from labs whose work depends on closed, proprietary systems: their techniques are inspectable, reproducible, and remixable by other researchers. For the broader AI community, that's valuable even apart from any single result — methods that others can build on compound faster than secrets.

Why a Japanese lab matters

Sakana AI is part of a broader story: serious frontier AI research happening outside the US-China axis. Japan brings deep strengths in robotics, manufacturing, and engineering culture to the AI race, and Sakana's emergence — backed by significant funding — signaled that Tokyo intends to be a player, not a spectator. Their focus on efficiency over scale also resonates in a country where computational resources are scarcer than in Silicon Valley.

The founders' pedigree

Sakana's credibility starts with its founders. Llion Jones co-authored "Attention Is All You Need" (2017), the paper that introduced transformers — the architecture behind virtually every modern AI system. David Ha spent years at Google Brain working on generative models and artificial evolution, including well-known research on world models and evolving neural networks. This isn't a lab founded on hype; it's founded by people who helped build the field's foundations and then chose to question its direction. That contrarian streak — insiders deliberately working against the scaling consensus — is what makes their output worth watching even when you disagree with the bets.

Fugu: the research bet now has a product

For years, everything on this page described a lab and its ideas. That changed in 2026: Sakana's multi-agent orchestration system, Fugu, is a real, shipping product. It entered beta in April, reached general availability in June, and on September 11, 2026 Sakana launched Fugu Max v1.0 and Fugu Ultra v2.0 — the two flavors the product line now revolves around.

The core idea will sound familiar if you've read the sections above. Fugu isn't a single foundation model; it's an orchestrator. You send one request through an OpenAI-compatible API, and behind the scenes it breaks the job into parts, routes each part to the best model in a swappable pool of open-weight and specialist models, and stitches the results into one answer. It's the evolutionary-coordination philosophy from the lab's research, now running as infrastructure. Our explainer on what Fugu is and how the orchestration actually works walks through the TRINITY and Conductor research papers behind it, in plain language.

The two tiers split on a simple question. Fugu Max ($2 input / $6 output per million tokens) asks "what's the best output per dollar?" — built heavily on open models including NVIDIA's Nemotron family, and aimed at multi-step agentic work where orchestration has something to orchestrate. Fugu Ultra v2 ($5 input / $30 output per million, with cached input at $0.50) asks "what's the highest capability on hard, multi-step tasks?" — and posts headline benchmark numbers without any frontier model in its pool. Our Fugu Max vs Ultra pricing breakdown covers who each tier is actually for.

And then there's the bigger story — the reason this launch got covered as more than a product announcement. A system that hits frontier-level results by coordinating open models changes the vendor lock-in equation (no single model in the pool is load-bearing) and complicates the export-control picture. We dig into why Fugu matters beyond the benchmarks: lock-in, supply-chain resilience, and what it means when the moat is the conductor, not the model.

How to follow their work (without a PhD)

Sakana publishes like an academic lab: papers, blog posts, and code. If you're curious but not a researcher, start with their blog posts, which usually explain each project in plain language with visual examples. The papers themselves are worth skimming for the introduction and conclusion even if the math is opaque — that's where the "why this matters" lives. And watch for the community response: when a Sakana release gets independent replication or spirited rebuttals, that's the signal it's touched something real. Hype fades in weeks; ideas that survive contact with other researchers are the ones that matter.

What to watch

• Model merging and evolution research — techniques for building better models from existing ones rather than training from scratch.
• Automated science — how far the AI Scientist concept goes, and how the research community responds.
• Japan's AI ecosystem — Sakana as the flagship of a growing domestic scene.

Following their work

Sakana AI publishes research papers and announcements through their official channels. The lab's work is technical and research-grade — aimed at the AI research community rather than end users.

In short: Sakana AI is a real lab, doing real research, in a direction most of the industry isn't exploring. Whatever you think of the bets, a field with only one strategy is a fragile field — and Sakana is the most credible team working on the alternatives.