Sakana AI Launches Fugu, an Orchestration Model Rivaling Fable 5

On June 22, 2026, Sakana AI launched Fugu. The model does not try to out-train the giants. Instead, it orchestrates them. The Tokyo lab is now Japan’s most valuable AI startup. Its new system routes your request across a pool of expert models, then hands back one answer. Below, we unpack what Fugu is, how it compares to Anthropic’s Fable 5, and the story behind the company.

So what does Sakana AI do? In short, it builds frontier AI with a different playbook. The Tokyo lab launched in late 2023. Its name means ‘fish’ in Japanese — a nod to the swarms and collective intelligence that inspire its research. Moreover, the company avoids pouring everything into one giant model. Instead, it favors nature-inspired ideas such as evolutionary model merging, which breeds strong models into a better one. That contrarian streak runs through everything it ships, and Fugu puts the philosophy on full display.

The biggest Sakana AI news today is Fugu, released on June 22, 2026. However, Fugu is not a single large language model. It is an orchestration system. One model learns to call on a pool of other AIs. It decides who handles what, then merges their work into one reply. Indeed, it ships in two flavors. A fast, low-latency Fugu handles everyday tasks. By contrast, a heavier Fugu Ultra tackles complex, multi-step problems like code review, data science, and paper reproduction. Crucially, both sit behind one OpenAI-compatible API. To the user, the whole pool behaves like a single model.

Under the hood, Fugu is a language model that has learned when to answer directly and when to delegate. In fact, it sometimes calls copies of itself recursively. The approach grew out of two Sakana papers at ICLR 2026, Trinity and Conductor, on learned model orchestration. The underlying pool is swappable. So Fugu can cut your dependence on any single provider. That matters after recent export controls limited access to Anthropic’s models. In effect, the lab is betting that smart coordination can rival raw scale — the same instinct that shaped its earlier research.

The numbers are the headline. According to The Decoder, Fugu Ultra posts 73.7 on SWE-Bench Pro. That beats GPT-5.5 (58.6), Opus 4.8 (69.2) and Gemini 3.1 Pro (54.2). It also scores 95.5 on GPQA-D and 93.6 on MRCRv2. That said, Sakana frames Fugu Ultra as standing shoulder-to-shoulder with Fable 5 and Mythos, not beating them outright. On a handful of tests, Fable 5 still leads. Even so, matching the frontier without training one is striking. It lands Fugu among the strongest best AI coding tools for code review and debugging.

A surprising number of people ask whether Sakana AI is based in San Francisco. It is not. Instead, the company sits in Tokyo, Japan. That location anchors its mission: building sovereign, world-class AI for a country long reliant on American models. The confusion is understandable, because the founders are Silicon Valley veterans. Still, the company runs its labs, government ties, and banking partnerships in Japan. That is exactly why investors treat it as a national champion.

The Sakana AI founders read like an AI all-star list. For instance, CEO David Ha previously led research at Stability AI and worked at Google Brain. Earlier, he built a career at Goldman Sachs. Meanwhile, CTO Llion Jones co-authored the landmark 2017 paper ‘Attention Is All You Need.’ That work introduced the Transformer, which underpins nearly every modern chatbot. COO Ren Ito, a former diplomat, rounds out the leadership. Together they walked away from Big Tech to prove a sharper lab could compete. Fugu is their clearest argument yet.

The money tells its own story. A $135 million Series B in November 2025 lifted the Sakana AI valuation to roughly $2.65 billion. That made it Japan’s most valuable unlisted startup and a clear unicorn. Earlier rounds drew an unusual roster of backers. For example, they include NVIDIA, Japanese banking giant MUFG, and even In-Q-Tel, the venture arm tied to the CIA. Many readers also ask about Sakana AI stock. There is none yet. After all, the company is privately held, so there is no ticker and no IPO on the calendar. For now, the only way in is as a partner or a customer.

Fugu did not appear from nowhere. Sakana AI first drew global attention with The AI Scientist. That system automates research end to end, from hypothesis to written paper. A v2 followed, not without controversy — an early version edited its own runtime code. The lab also pioneered evolutionary model merging and explored ideas like its Continuous Thought Machine. Fugu applies that same instinct to inference: do not build one brain, coordinate many. Ultimately, it is a distinctly local answer to a problem rivals attack with sheer size.

Fugu is the latest shot in a fast-moving model race. For the rival it challenges, read our breakdown of Claude Fable 5, Anthropic’s most powerful public model. And for another challenger undercutting the incumbents on price and code, see how GLM 5.2 is shaking up the benchmarks.

*Sources: TechCrunch, GIGAZINE, Startup Fortune, Sakana’s official Fugu release.*