Japan Couldn’t Win the AI Arms Race. So Sakana AI Changed What the Race Looks Like.

A Tokyo startup founded by a co-author of the Transformer paper just matched Fable 5’s benchmark scores using a 7-billion parameter conductor and other companies’ models. The timing, four days after the US export ban, was not accidental.

There is a sentence in Sakana AI’s Series B announcement from late 2025 that has aged unusually well.

“We believe that intelligent life has arisen not from an abundance of resources but rather from the lack of it.”

Sakana AI is a Tokyo-based company. Japan does not have the compute infrastructure of the United States. It does not have the data center buildout of the Gulf. It does not have the state-sponsored capital that has fueled Chinese AI development. It has roughly a third of the GPU capacity per researcher that major American labs operate with, limited domestic training data relative to English and Chinese internet corpora, and a declining workforce that makes the talent problem acute.

These are not framing devices. They are real constraints that shaped a real strategic decision. Sakana AI decided several years ago not to compete in the race to build the largest monolithic model. It decided instead to figure out how to make AI systems smarter by having them work together.

On June 22, 2026, that bet produced its most visible result yet: Sakana Fugu, a multi-agent orchestration system that claims performance comparable to Anthropic’s Fable and Mythos models on certain benchmarks.

The timing, four days after the US Commerce Department ordered Anthropic to pull both of those models from global access, was not accidental.

Who Built This

Before getting to the architecture and the benchmarks, the founding team is worth understanding, because it explains the intellectual ambition of what Sakana is attempting.

Llion Jones is one of eight co-authors on the 2017 paper “Attention Is All You Need,” the paper that introduced the Transformer architecture that underlies essentially every major AI model deployed at scale today. He left Google and moved to Tokyo to co-found Sakana in 2023, alongside David Ha, a former Google Brain researcher who had led the AI team at Google Japan and published significant work on evolutionary computation and collective intelligence.

That background shows in what Sakana has chosen to research. The company is not trying to out-GPT GPT. It is exploring the questions that sit at the boundary between the current paradigm and whatever comes next. Their published work includes the Evolutionary Model Merge method, which fuses capabilities from existing open-source models through evolutionary algorithms. Their Continuous Thought Machine paper proposed a new architecture inspired by temporal processing in the brain, a different structure than the Transformer that Jones helped invent. The AI Scientist, one of their most-discussed systems, is a multi-agent framework that autonomously handles research tasks from hypothesis generation through experiment running to paper writing.

These are not the research priorities of a lab trying to win the benchmark race. They are the research priorities of a lab that thinks the benchmark race is not the interesting question.

Fugu is where that research philosophy met a product decision.

What Fugu Actually Is

The description that Sakana uses is precise and worth quoting directly: a multi-agent system that behaves like a single model.

The operational reality behind that description is this. When you send a request to Fugu, you send it to one endpoint. You interact with it through an OpenAI-compatible API, meaning you can switch to it from a standard OpenAI client with minimal code changes. But inside that endpoint, Fugu is not a single large model generating a response. It is a coordinator deciding how to handle your request.

The coordinator is the key architectural component: a language model of about 7 billion parameters, trained through reinforcement learning specifically to understand how to delegate tasks across a pool of frontier models. Not trained to answer questions well. Trained to understand when a question should be answered directly, when it should be broken into subtasks and distributed, which subtasks should go to which models, how to verify the outputs, and how to synthesize them into a coherent response.

The research behind this approach comes from two ICLR 2026 papers: Trinity, which studies evolved LLM coordination, and the Conductor, which addresses learning to orchestrate agents in natural language. The central finding from this research is that a small model trained specifically on coordination tasks can outperform what any individual model in the pool achieves alone. The orchestration itself adds value beyond the sum of the parts.

This is architecturally different from what most people mean when they say “multi-agent system.” Most multi-agent architectures, including the kinds you can build with LangGraph or AutoGen, use fixed workflows that a human engineer defines in advance. Orchestrator agent calls tool A, then tool B, checks the output against criteria C, escalates to human if D occurs. The routing logic is code.

What Fugu does instead is learn the routing. The conductor model figures out delegation strategies from experience. Because Fugu is built on learned orchestration rather than fixed workflows, it improves as the underlying ecosystem improves: as new frontier models arrive, they can be folded into the pool. The system does not need to be redesigned when a better model becomes available. It incorporates it.

At launch, Sakana has not disclosed the full list of models in Fugu’s pool. GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8 are confirmed. The system also supports calling copies of itself recursively for tasks that benefit from nested orchestration. Two variants ship: standard Fugu, optimized for everyday coding, review, and chat with lower latency, and Fugu Ultra, which coordinates a deeper pool of expert agents and is oriented toward complex, multi-step, high-stakes tasks. For Fugu standard, users can opt specific models out of the pool, which matters for teams with data governance requirements. For Fugu Ultra, the pool is fixed.

The Numbers and What They Actually Mean

On SWE-Bench Pro, Fugu Ultra scored 73.7. Claude Opus 4.8 scored 69.2. GPT-5.5 scored 58.6. On TerminalBench 2.1, Fugu Ultra scored 82.1, compared to GPT-5.5’s 78.2 and Opus 4.8’s 74.6.

These are not cherry-picked easy benchmarks. SWE-Bench Pro tests autonomous resolution of real GitHub issues against production codebases. TerminalBench tests system operation and tool-calling stability in multi-step, long-chain tasks. Beating the individual models it is coordinating on these specific benchmarks is the cleanest demonstration of the orchestration value-add, because it shows the conductor is doing something beyond simple routing.

Sakana also claims Fugu Ultra achieved scores comparable to Anthropic’s Mythos Preview and Fable 5 on GPQA-Diamond, CharXiv Reasoning, and Terminal Bench. The company acknowledges it did not outperform those models on every test, but argues the results suggest that coordinating existing frontier models can sometimes rival the performance of a single cutting-edge system.

The honest framing is important here. Fugu Ultra did not beat Fable 5 overall. On the specific benchmarks Sakana published, it matched or came close. On benchmarks Sakana did not publish prominently, we do not know. Ethan Mollick, an associate professor at Wharton who tests frontier models extensively, reported that in practice the system is incredibly slow, with his typical coding tests taking 30 minutes to run, and concluded that the results are fine but do not match Fable in real use.

Latency is not a trivial objection. A system that produces excellent outputs in 30 minutes competes with a different category of use cases than a system that produces good outputs in 30 seconds. For tasks where you are running a long autonomous research workflow overnight and care only about final output quality, the latency may be acceptable. For interactive coding, where you are expecting a suggestion within a few seconds, 30 minutes is not a usable product.

Early public reaction skewed skeptical, with a Sakana-commissioned review of posts finding 3 supportive, 6 skeptical, and 3 critical. The dominant objection on Hacker News and technical Twitter in the first 24 hours was a version of: is this a new kind of AI or an expensive router that happens to use good models? The “is this just a wrapper?” framing appeared repeatedly.

That objection is partially fair and partially confused. Calling Fugu a wrapper misses what the learned orchestration component actually does. A fixed router calls GPT-4 for some requests and Claude for others based on predefined rules. Fugu’s conductor learns which combination of delegations, verifications, and syntheses produces better results for which task structure. That is a meaningful technical contribution. At the same time, Fugu’s outputs are fundamentally dependent on the underlying models in its pool. If those models go offline, get export-controlled, or change their behavior through fine-tuning, Fugu’s performance changes. That is a real dependency that a monolithic model does not have.

Aaron Levie, CEO of Box, made the most balanced public observation on launch day: this is generally how sophisticated applied AI products are already building their agent harnesses, but making it available as a single API that any developer can call is a genuinely useful abstraction, and the routing layer will produce more value as both frontier closed and open-source models continue to improve.

The Timing Was Not a Coincidence

Sakana AI explicitly cited the export controls on Anthropic’s Fable and Mythos models as motivation for Fugu’s design. The framing in their documentation describes the multi-provider pool as a hedge against single-vendor dependency: if one provider restricts access, Fugu routes around the disruption.

This is a smart and direct response to a real problem that became dramatically more concrete four days before Fugu’s launch.

When Anthropic’s Fable 5 and Mythos 5 disappeared behind export controls on June 12, every enterprise and government organization outside the United States that had integrated those models received a sudden, involuntary demonstration of what single-vendor AI dependency looks like in practice. A model they had deployed into production workflows returned a 404 error. Their options were limited: wait for access to restore, or rebuild around something else.

Fugu’s architecture is explicitly designed to prevent that outcome. With multiple models in the pool, including both American and non-American frontier models, no single export control action takes the entire system down. If Anthropic’s models become unavailable, the conductor routes around them. If OpenAI faces a regulatory action, the same. The resilience is architectural rather than requiring manual intervention.

This is a meaningful competitive position, and it is not available to any single-model provider. You cannot make Fable 5 resilient against its own export control by making it better at coding. You can only make it resilient against export controls by not depending on it exclusively. Fugu’s design operationalizes that insight.

The framing that Sakana uses publicly, AI sovereignty through orchestration, deliberately echoes the language that European governments have been using about reducing dependency on American AI infrastructure. By orchestrating the world’s models, they describe themselves as delivering the realistic, resilient blueprint required for AI sovereignty. This is positioning as much as it is technical description, but the underlying claim is not wrong. An organization running Fugu can achieve reasonable continuity of AI service in the face of geopolitical disruptions that would take down any single-provider deployment.

The Japan Strategy

The most interesting thing about Sakana is not any single technical result. It is the strategic clarity with which they have defined what Japan’s AI development should look like given Japan’s actual situation.

Sakana has argued that Japan, with its limited resources, declining workforce, and aging population, should focus on post-training optimization and model orchestration rather than trying to compete in large-scale pretraining against labs with far larger compute and data resources. They have built enterprise partnerships with MUFG and Daiwa Securities Group for custom financial AI, expanding into defense and manufacturing. They received 32 billion yen in their Series B, about $200 million US dollars, at a valuation of over $2.5 billion.

This is a sustainable business built around a realistic assessment of comparative advantage. Japan has deep domain expertise in manufacturing, finance, and engineering. It has a cultural and organizational tradition of precision, reliability, and long-horizon thinking. What it does not have is the ability to out-compute the United States or out-data China in general-purpose pretraining races.

So Sakana is not trying to do that. They are trying to be the best in the world at figuring out how to get the most out of the models that already exist, and at customizing those models and orchestration layers for specific high-value Japanese industry applications.

Fugu is part of that strategy, but only part. The research into nature-inspired AI, evolutionary algorithms, and architectural alternatives to the Transformer is another part. The enterprise partnerships are another. The company is building multiple options in a world where no single bet on any AI approach is clearly correct.

What This Architecture Actually Changes

The “is it just a router?” objection undersells the contribution but also points to something real.

What learned orchestration changes, if it works at scale, is the relationship between model capability and system capability. Under the current paradigm, when you need more capability, you wait for the next frontier model. You are dependent on a single lab’s development cycle, safety review process, and pricing decisions. When you get the new model, you integrate it and benefit.

Under an orchestration paradigm, when a new model becomes available, you add it to the pool and the coordinator learns how to use it. Your system improves without you redesigning anything. More importantly, you can combine capabilities that no single model has. A model with exceptional coding ability but mediocre scientific reasoning can be paired with a model that has the opposite profile. The coordinator learns which problems to route where.

This does not eliminate the value of frontier model development. Fugu needs good models in its pool. Better models in the pool should produce better Fugu outputs. The two approaches are not in competition. But it does mean that the output quality ceiling is not set by any single model’s capability. It is set by how well the coordinator can combine what is available.

Whether Sakana’s specific implementation of this idea reaches the potential the theory suggests is what the next twelve to eighteen months of real-world deployment will reveal. The benchmark scores are encouraging. The latency problem is real and needs to be addressed before the system is competitive for interactive use cases. The dependency on third-party frontier models is a genuine architectural vulnerability even as it is also a design advantage in geopolitical resilience terms.

But the core idea, that a small model specifically trained to coordinate larger models can outperform any individual model in the pool, has now been demonstrated on real benchmarks. The 7-billion parameter RL Conductor is not the interesting part in isolation. What is interesting is that it beats the 400-billion parameter models it is calling.

That result, if it holds and extends, changes something about how the field thinks about where capability comes from. Parameters have been the primary currency of AI development for a decade. Fugu is an argument that coordination can be a different currency, and that you can achieve effects well above your parameter weight class if you know how to assemble the right team.

Sakana was founded by someone who helped invent the architecture that made the parameters race possible. It is fitting that they are now building the system that might make the race less one-dimensional.