.svg){kind=spacer}
Two weeks. That is roughly how long Sakana AI says it now needs to fold a newly released frontier model, a Gemini, a GPT, a Claude, into its own system and pass the gains on to customers. Not two years of training. Not a nine-figure compute run. Two weeks to absorb someone else's breakthrough and turn it into your own edge.
On June 22, 2026, the Tokyo lab released Sakana Fugu, and with it a claim that should make every founder in Asia sit up: its top model, Fugu Ultra, stands shoulder-to-shoulder with the most capable systems on the market across the industry's hardest engineering, science, and reasoning benchmarks, without Sakana having trained a frontier model at all. On SWE-Bench Pro, a test of whether an AI can fix real software bugs, Sakana reports Fugu Ultra at 73.7%, ahead of Anthropic's Claude Opus 4.8 (69.2%), OpenAI's GPT-5.5 (58.6%), and Google's Gemini 3.1 Pro (54.2%). Across Sakana's own comparison table, Fugu Ultra tops ten of eleven benchmarks, as the independent explainer site DataCamp noted in its coverage.
Here is the twist that matters. Fugu did not beat those models. Fugu used them. It is what Sakana calls a multi-agent system delivered as a single model, an orchestrator trained to route each task to the best expert in a pool that includes GPT-5.5, Opus 4.8, and Gemini 3.1 Pro, and to make them collaborate. The tagline Sakana chose says it plainly: "One Model to Command Them All." The frontier, in other words, was reached not by building a bigger brain, but by hiring a better conductor.
That reframing is worth far more to Asia's builders than one product launch. For a decade, the unspoken rule of the AI race has been that the winners are whoever can afford the largest training runs, a contest of capital and compute that Southeast Asian founders, and most of the world, were never going to win. Sakana's bet is that the next phase rewards a different skill entirely: not who trains the biggest model, but who best coordinates everyone else's. If that bet is right, an arms race Asia could never win becomes a coordination game it can.
This is the story of how a small Japanese lab got there, and the harder, more contested terrain that a strategy built on orchestrating other people's models must still cross.
To see why Fugu matters, start with the strategy it is quietly abandoning. For the past several years, progress in artificial intelligence has been driven by brute-force scale: bigger models, more parameters, ever-larger training runs. It worked spectacularly, and it produced a natural gatekeeper. Training a frontier model now demands compute budgets only a handful of American labs and their cloud backers can command, which is precisely why the benefits of frontier AI have concentrated in the hands of those few.
But two cracks have opened in the monolithic model. The first is that raw scale is hitting diminishing returns, and the second, more useful, is that frontier models have started to specialise. In Sakana's technical report, the pattern is documented in detail: GPT-series models tend to shine at mathematical reasoning, Opus-series models at software engineering and cybersecurity, Gemini at certain scientific and factual-recall tasks. No single model is best at everything. The report frames the opportunity in one line worth keeping: the next frontier may be reached not by any single model, but by a system that can identify, combine, and amplify the complementary strengths of many.
That is the intellectual bet Sakana has been building toward since its founding. The lab was co-founded by Llion Jones, one of the authors of the 2017 paper "Attention Is All You Need," which gave the world the transformer, and David Ha, and it has argued from the start that the most powerful AI systems will be collaborative ecosystems, not isolated monoliths. The name itself is the thesis in miniature: sakana means fish in Japanese, and the company keeps returning to the image of a school of fish that behaves, through coordination, as a single intelligent organism.
Fugu is that idea, productised. Under the hood, as Sakana details in its report, the system is trained in a way no traditional multi-agent framework attempts. The lighter model, simply called Fugu, learns to pick the single best model for each request through a lightweight selection head, refined first by large-scale fine-tuning and then by evolutionary optimisation, a nod to Sakana's earlier work using evolution to merge AI models. The heavier model, Fugu Ultra, builds on the lab's reinforcement-learning research to design entire workflows on the fly, assigning up to five steps across a team of expert agents, deciding who drafts, who checks, and who has the final word. Both are reached through a single API. From the outside you call one model; on the inside, a coordinated orchestra is doing the work.
The archetype of the old approach is the soloist: one enormous model, trained at enormous cost, asked to do everything alone. It carried AI a long way. But it left the frontier locked behind a compute budget, and that is a door orchestration was built to pick.
A strategy this clever invites hard questions, and the honest ones cut in two directions: whether Fugu delivers what it claims, and whether "orchestrating other people's models" is the independence it is sold as.
Every benchmark figure above comes from Sakana, and as DataCamp flagged plainly, these scores have not yet been independently reproduced by third-party labs. Impressive internal results are the starting gun of a credibility contest, not the finish line; until outside researchers rerun them, the right posture is interest, not conviction. Sakana deserves credit for publishing a detailed technical report and running a beta with close to 500 external users — but the verification that turns a claim into a fact still lies ahead.
Fugu decides which underlying model handles your request, and by design it will not tell you which one. Sakana treats the routing as proprietary. For a great deal of work that is fine; for regulated industries that must audit their reasoning chains, both DataCamp and the India-based community site Analytics Vidhya singled this opacity out as a real limitation. Coordination you cannot inspect is a hard sell to a compliance officer.
Routing, delegating, and synthesising across multiple models adds latency and overhead. For a simple query, a direct call to a single frontier model will usually be faster and cheaper — a trade-off both independent write-ups were careful to name. Fugu earns its keep on long, messy, multi-step problems; on quick ones it can be the expensive answer to an easy question.
Sakana frames Fugu explicitly as a hedge against a danger it argues is no longer hypothetical. In its launch note the lab pointed to the export controls that briefly cut off access to Anthropic's Fable and Mythos models in June 2026 as evidence that access to a single vendor's AI can shift or vanish overnight with a change in regulation — and argued that a swappable pool of models is the practical answer, routing around any one provider that disappears. It is a serious argument, and the operational logic holds: lose one model, and Fugu re-routes. (Those particular restrictions proved temporary — access was subsequently restored — but the broader point about fragility stands, which is exactly why the episode resonated.)
Yet notice what the hedge does not buy. Fugu's whole capability rests on the frontier models in its pool — GPT-5.5, Opus 4.8, Gemini 3.1 Pro, all built by the same handful of labs whose concentration prompted the worry. It is genuine insurance against losing any one provider. It is not independence from the frontier itself; it is a smarter way to depend on all of it at once. Whether that amounts to real "AI sovereignty," as DataCamp put it, is debatable — even as the engineering underneath is sound. An orchestrator is only ever as capable as the musicians it can hire.
The ceiling here, then, is not that orchestration fails. It is that orchestration inherits every strength and every dependency of the models it conducts — and asks us to trust a score we are not allowed to read.
Here is the reframing this moment offers Asia's founders. For a decade, the entry ticket to frontier AI was the ability to train a frontier model — a ticket priced in hundreds of millions of dollars and gigawatts of compute, and effectively unavailable outside a few firms. Orchestration changes the toll. If capability can be amplified by composing existing models rather than only by training larger ones, then, as Sakana argues in its report, progress need not depend solely on access to the largest training runs — and its benefits can be distributed more broadly across organisations and regions rather than concentrated in those able to build the biggest models. That single sentence is the opening. Three shifts follow from it.
The most valuable position in the old game was the one almost no one in Asia could occupy: builder of the base model. Fugu points to a different high ground — the orchestration layer that sits above the models and decides how to use them. That layer is reachable without a training-scale war chest, and it is where a great deal of real-world value now lives. For founders, the practical read is that you do not have to out-train OpenAI to build something that outperforms any single one of its models on your task; you have to out-*coordinate*.
Most teams building agentic products today stitch together their own orchestration by hand, wiring up frameworks like LangGraph, AutoGen, or CrewAI and then maintaining the whole fragile apparatus. Fugu's proposition, as Analytics Vidhya's walkthrough emphasised, is that this coordination becomes a property of the model itself, reached through an OpenAI-compatible endpoint with no SDK migration — and, for the lighter model, with the ability to opt specific providers out of the pool to meet data and compliance needs. The lesson is not "use Fugu"; it is that the coordination problem, not the raw model, is increasingly where the engineering leverage sits.
The most quietly radical result in Sakana's report is not a coding benchmark at all. It is a test of whether an AI can recover the reading order of classical Japanese kana letters written in the chirashigaki, or "scattered-writing," style — characters strewn across the page in a way that even trained scholars find hard to parse, and for which no training data exists and, the report notes, cannot readily be made to exist. On a corpus of twenty-five expert-annotated letters, Fugu Ultra scored a mean 0.776 by the study's edit-distance measure, ahead of the strongest frontier baseline at 0.642, while another leading model could not complete the task at all. It is a deliberately, unmistakably Japanese problem — and Fugu, by orchestrating models toward it, solved what none of them could solve alone. That is what the frontier looks like from the ground: not a foreign template localised, but a local problem that forces a new capability into being.
The smart money and the smart builders are already moving toward this layer. Sakana's system is available not only through its own console but via third-party gateways such as OpenRouter and Vercel's AI gateway, and the lab has committed to folding each new frontier model into Fugu's pool within roughly two weeks of its public release — a design in which the system improves automatically as the whole ecosystem improves. The coordinators, in other words, compound. While the crowd watches the next big model, the interesting question has quietly become who conducts them best.
For the founders Asia Tomorrow exists to serve, Sakana's gambit distils into a handful of hard, usable truths.
You will not out-train the frontier labs, and you no longer have to. The defensible ground is the coordination and application layer above the models — reachable without a training-scale budget, and increasingly where value accrues. Build there.
Sakana's most practical insight is architectural: never let a single provider become a single point of failure. Design so that if one model's access, price, or policy shifts overnight, your product re-routes and survives. That discipline is worth more the more the geopolitics of AI tighten.
Fugu's most valuable demonstration was a Japanese task with no training data and no Western template. The durable edge for Asian founders sits in exactly these places — local languages, local systems, local constraints — where there is nothing to import and localise, only something to invent.
Orchestration's opacity is a genuine limitation, not a detail. If your work must be audited, weigh what you cannot see — and let that shape which layer of the stack you build on and how.
Self-reported benchmarks, however striking, are the beginning of a credibility contest. Reserve conviction for what independent hands can reproduce — and hold your own claims to the same bar.
Conclusion: The World Needs More Conductors
Asia's place in the AI story has too often been cast as the follower — the region that adopts what others invent, one generation late and one compute budget short. Sakana Fugu is a small, pointed rebuttal. A Tokyo lab, without building a frontier model, has produced a system that stands with the frontier by orchestrating it — and in doing so has sketched a path to the frontier that does not run through a nine-figure training run.
The claims still need outside verification, the black box still needs opening, and the sovereignty it promises is really a smarter form of dependence rather than escape from it. Those are real limits, and honest builders should hold them in view. But the reframing underneath is the part worth carrying: in the next phase of AI, capability may belong less to whoever owns the biggest model and more to whoever best coordinates the many — a game measured in ingenuity rather than gigawatts, and one in which Asia's founders are not spectators but contenders.
The soloists were always going to be few, and far away, and expensive. The orchestra is open to anyone who can learn to conduct. That is why the world needs more of Asia's builders reaching for the baton — not fewer.
Sources
1. Sakana AI (2026), "Sakana Fugu: One Model to Command Them All," launch announcement, 22 June 2026. https://sakana.ai/fugu-release/ — Primary source for the launch date and framing, the description of Fugu as a multi-agent system delivered as a single model, the two-variant (Fugu / Fugu Ultra) structure, the ~500-user beta, the roughly two-week cadence for adding new frontier models, the beta user reports, and Sakana's own "AI sovereignty" / export-control argument (which links to Anthropic's statement on Fable and Mythos access).
2. Sakana AI, "Sakana Fugu — Multi-Agent System as a Model," product page (accessed July 2026). https://sakana.ai/fugu/ — Source for the full benchmark table (Fugu and Fugu Ultra vs Opus 4.8, Gemini 3.1 Pro, and GPT-5.5), the pricing (Fugu Ultra at $5 input / $30 output per 1M tokens; subscription tiers at $20 / $100 / $200), the qualitative demonstrations (AutoResearch, classical-kana reading order, Rubik's-cube solver, CAD mechanical iris, blindfold chess, stock trading), third-party availability (OpenRouter, Vercel AI gateway, opencode), the agent opt-out feature, and the EU/EEA availability restriction.
3. Fugu Team, Sakana AI (2026), *Sakana Fugu Technical Report*, arXiv:2606.21228 (cs.LG), 19 June 2026. https://arxiv.org/abs/2606.21228 — Primary technical source for the orchestration architecture, the training methods (large-scale fine-tuning, evolutionary strategies for Fugu, reinforcement learning / the Conductor and Trinity frameworks for Fugu Ultra), the framing of orchestration as a complementary scaling axis, the documented model specialisations, the composition of the agent pool (Gemini 3.1 Pro, Claude Opus 4.8, GPT-5.5), the AutoResearch result (123 experiments, ~14 hours on a single H100), and the classical-kana reading-order study (mean edit-distance scores across 25 expert-annotated letters). Licensed CC BY 4.0; project lead Yujin Tang.
4. DataCamp, "Sakana Fugu: Features, Benchmarks, and How It Works," 24 June 2026. https://www.datacamp.com/blog/sakana-fugu — Independent analysis used for the SWE-Bench Pro comparison and the "tops 10 of 11 benchmarks" observation, the Llion Jones / David Ha founding note, and — importantly — the critical caveats: that the benchmark numbers are Sakana-reported and not yet independently reproduced, that the routing opacity is a limitation for compliance-sensitive work, that the latency/cost trade-off is real, and that the "AI sovereignty" characterisation is debatable.
5. Harsh Mishra, "Sakana Fugu: Multi-Agent System as a Model," Analytics Vidhya, 23–24 June 2026. https://www.analyticsvidhya.com/blog/2026/06/sakana-fugu-multi-agent-system-as-a-model/ — Independent (India-based) analysis used for the "school of fish" collective-intelligence framing and the meaning of the name, the positioning against hand-built orchestration frameworks (LangGraph, AutoGen, CrewAI), the OpenAI-compatible integration and one-to-three-agent routing for Ultra, and a second confirmation of the transparency, latency, token-accounting, and regional caveats.
6. Anthropic, statement on Fable and Mythos model access (2026). https://www.anthropic.com/news/fable-mythos-access — Cited by Sakana as the concrete instance of a vendor's model access shifting under export controls; referenced here for the June 2026 access episode that Sakana uses to motivate its single-vendor-dependency argument. (Model-availability specifics in this area move quickly and were still evolving at the time of writing.)
7. Sakana AI, Sakana Fugu code and API repository. https://github.com/SakanaAI/fugu/ — The public repository accompanying the OpenAI-compatible API; noted for developers seeking the client and integration details. (Referenced as a resource rather than mined for claims in this piece.)
0 Comments