Japan has not suddenly discovered artificial intelligence in 2026. The country already has decades of strength in robotics, semiconductors, industrial automation and machine learning, and companies such as SoftBank, NTT, Preferred Networks and NEC have been building AI systems for years.
What has changed is the shape of Japan’s frontier-AI ambition.
Sakana AI, a Tokyo-based startup founded in 2023, has built Fugu, a system that does not try to win the AI race by training one gigantic model and asking it to do everything. Instead, Fugu acts as an intelligent coordinator. It receives a task, decides whether one model is enough, and when the problem is harder, it can assemble multiple AI agents, assign work, coordinate their interaction, verify intermediate results and synthesize a final answer through a single API.
That makes Fugu less like another chatbot and more like a control layer for artificial intelligence.
Sakana calls it “a multi-agent system delivered as one model.” The phrase matters because the product hides the complexity of model selection and collaboration behind an interface that developers can call like a normal foundation model.
The broader strategic idea is even more important.
OpenAI, Anthropic, Google and other frontier labs have spent enormous sums scaling individual models. Sakana is betting that some future capability gains can come from better coordination rather than simply building a larger neural network.
If that thesis works, Japan may have found a distinctive route into the frontier AI race: not by copying the scale economics of the United States or China, but by building an orchestration layer that can combine the strengths of models from multiple providers.
What is Sakana Fugu?
Sakana Fugu is a family of orchestrator models developed by Sakana AI.
It is itself a language model, but its main job is not simply to generate the answer token by token. Its job is to decide how a problem should be solved.
For a relatively simple request, Fugu can choose a direct path.
For a difficult multi-step problem, it can dynamically coordinate several agents backed by frontier models. It can decide which model should plan, which should execute, which should verify, and how information should flow between them.
From the developer’s perspective, this still looks like one model call.
That abstraction is the product.
Traditional multi-agent systems can require developers to manually create roles, define routing logic, build retry rules, manage model-specific APIs and decide when one agent should pass work to another.
Fugu attempts to learn that coordination behavior.
Its technical report describes the system as dynamically constructing “agentic scaffolds” based on the user’s query.
That means the workflow itself can change depending on the problem.
Fugu is not simply a router
Calling Fugu a model router understates the ambition.
A basic router might inspect a prompt and send coding questions to one model, mathematics to another and writing tasks to a third.
Fugu can go further.
Sakana’s research explores systems in which multiple models collaborate over several turns, take different roles, critique one another and revise results.
The orchestrator therefore has to answer several questions:
Which agents should be involved?
How many are necessary?
What instructions should each receive?
In what order should they work?
Should one agent verify another?
Should the system ask for another attempt?
When is the answer good enough to stop?
Those decisions are themselves an intelligence problem.
Sakana’s thesis is that learning to coordinate models can become a new scaling axis for AI.
The research behind Fugu: TRINITY
One of Fugu’s foundations is a research system called TRINITY.
The work was accepted at ICLR 2026.
TRINITY uses a compact coordinator to delegate among a pool of language models.
It can assign roles such as Thinker, Worker and Verifier.
The Thinker can reason about the problem and propose an approach.
The Worker can execute the task.
The Verifier can inspect the result and identify mistakes.
The key point is that those roles do not have to be permanently attached to one model.
The coordinator can decide which underlying model is best suited to a role at a particular point in the task.
This creates a kind of computational division of labor.
Human organizations do this naturally. A lawyer, engineer and accountant do not all solve a complicated transaction in the same way. One person may draft, another calculate and another audit.
TRINITY tries to make that kind of specialization adaptive inside an AI system.
The Conductor: learning how agents should communicate
The second major research foundation is Sakana’s Conductor work, also accepted at ICLR 2026.
Instead of relying on a human engineer to define the communication graph between agents, the Conductor learns how agents should exchange information.
It can generate targeted instructions for individual workers and determine communication patterns that help the overall system solve a task.
That is a subtle but important difference from many early multi-agent frameworks.
A hand-built agent framework may say:
Researcher sends findings to Writer.
Writer sends draft to Reviewer.
Reviewer sends corrections back to Writer.
That workflow is fixed before the task begins.
A learned conductor can potentially decide that a mathematics problem needs one topology, a coding task another and a long research problem a third.
Fugu commercializes this broader idea.
Why this approach is different from simply making a bigger model
The mainstream frontier-model strategy has been dominated by scale.
Train on more data.
Use more compute.
Increase model capacity.
Spend more at inference time.
Those methods have produced extraordinary gains.
But they are expensive.
They also assume that one model should contain or produce nearly all the capability a user needs.
Sakana is testing another idea: intelligence can emerge from coordination among specialists.
In this model, the orchestrator itself does not need to be the strongest coder, mathematician or scientist.
It needs to know when to call stronger specialists and how to combine them.
That resembles operating systems and computer architecture.
A CPU does not perform every function inside a modern computer. Different processors and accelerators handle different workloads while software coordinates the system.
Fugu is attempting something similar at the model layer.
Why this matters for Japan
Japan is unlikely to win a brute-force AI spending contest against the largest U.S. hyperscalers.
Alphabet, Microsoft, Amazon and Meta can each invest tens or hundreds of billions of dollars into data centers and accelerator fleets.
China is also mobilizing enormous public and private resources into AI infrastructure.
Japan needs a different strategic advantage.
Orchestration can potentially provide one.
If the best model for one task comes from an American provider, another from a Japanese lab and another from an open-source project, an orchestration system can combine them without requiring Japan to reproduce every frontier model internally.
That does not create complete technological independence.
The system can still depend on foreign models, cloud infrastructure and chips.
But it can reduce dependence on any single provider.
Sakana explicitly frames this as part of an AI-sovereignty strategy.
Fugu’s single-vendor argument
The geopolitical case became more important as access to frontier models became entangled with regulation and export controls.
Sakana argues that relying on one model provider creates operational risk for governments and enterprises.
A provider can change prices.
It can retire a model.
It can change its terms.
A country can introduce export restrictions.
A model can become unavailable in a particular jurisdiction.
Fugu’s architecture is designed around a swappable pool.
If one model becomes unavailable, Sakana can theoretically route around it and use another.
This is one of the most commercially interesting aspects of the product.
Enterprises do not like strategic dependency on one vendor.
A system that can preserve an application interface while changing the models underneath could reduce switching costs.
The challenge is whether performance remains stable when the pool changes.
That remains an engineering problem, not a solved law of nature.
The company behind Fugu
Sakana AI was founded in Tokyo in 2023 by David Ha, Llion Jones and Ren Ito.
Ha previously worked at Google and led research work connected to Google Brain in Japan.
Jones is one of the authors of the 2017 “Attention Is All You Need” paper that introduced the Transformer architecture underlying modern large language models.
Ito previously worked in Japan’s Ministry of Foreign Affairs and in technology businesses.
The combination is unusual.
Sakana has frontier AI research credentials, Japan-specific institutional knowledge and a corporate strategy centered on building a major research company in Tokyo rather than relocating to Silicon Valley.
That positioning has attracted substantial capital.
In November 2025, Sakana raised ¥20 billion, around $135 million at the time, in a Series B round that valued the company at approximately $2.65 billion post-money.
Its investor list includes major Japanese financial and industrial groups as well as global technology and venture investors.
Sakana became a unicorn unusually quickly
Japan’s Ministry of Economy, Trade and Industry has highlighted Sakana as one of the country’s fastest companies to reach unicorn status.
That matters because Japan historically has produced fewer venture-backed software giants than the United States or China.
The country has world-class incumbents in electronics, automotive manufacturing, robotics and industrial equipment, but its startup ecosystem has often been criticized for smaller funding rounds and a lower tolerance for extremely high-risk technology investment.
Sakana breaks that pattern.
Its valuation is not proof that Fugu will become a global platform.
Private valuations reflect expectations, not realized profit.
But the amount and quality of capital behind the company give it enough time to attempt technically difficult work.
June 2026 changed Sakana from research story to product story
Fugu moved from early beta into general availability in June 2026.
Sakana said nearly 500 early users had participated in its beta program.
The company launched two main variants.
Fugu is intended to balance performance and latency.
Fugu Ultra is designed for difficult tasks where maximum answer quality matters more than speed or cost.
Both are exposed through an OpenAI-compatible API.
That is strategically smart.
Developers already building against OpenAI-style APIs do not have to learn a completely new integration model to test Fugu.
The product can therefore compete at the interface layer even though its internal architecture is very different.
The benchmark numbers are impressive, but need discipline
Sakana reports strong results across coding, science and reasoning benchmarks.
For the original v1.0 release, Sakana reported Fugu Ultra at 73.7 on SWE-Bench Pro.
It reported 82.1 on TerminalBench 2.1.
LiveCodeBench was 93.2.
GPQA Diamond was 95.5.
Humanity’s Last Exam was 50.0.
On several of those tests, Sakana’s published results were above the baseline figures it presented for major frontier models.
These are important results.
They are not the same as independent proof that Fugu is universally better than GPT, Gemini or Claude.
Sakana’s own technical report says its baseline scores come from model-provider reports. The comparisons can therefore involve different provider evaluation setups.
The Fugu numbers are also vendor-reported.
Until independent labs reproduce results across the same harnesses and conditions, the responsible interpretation is that Fugu has demonstrated highly competitive benchmark performance in Sakana’s evaluation framework.
That is still a meaningful achievement for a young Japanese startup.
Fugu Ultra v1.1 shows the architectural advantage
In July, Sakana released Fugu Ultra v1.1.
The company said the updated system improved results by as much as 7.9 points across the benchmarks it tracks, with particularly strong gains in ProgramBench and TerminalBench 2.1.
The interesting part is how Sakana describes the improvement.
The company did not need to train one entirely new mega-model to make Fugu better.
It incorporated newer frontier models into the orchestration system and improved coordination.
This is central to the Fugu thesis.
A conventional foundation model can become obsolete when a stronger model is released.
An orchestrator can potentially benefit from the release.
If a new specialist enters the pool, the coordinator gains another capability.
That turns model competition into an input rather than only a threat.
Fugu’s business model also reveals its cost structure
Multi-agent intelligence is not free.
Every internal agent consumes tokens.
Sakana’s pricing makes this visible.
Fugu Ultra currently lists standard pricing at $5 per million input tokens and $30 per million output tokens, with cached input priced at $0.50 per million.
Long contexts above 272,000 tokens carry higher rates.
Sakana also warns that orchestration tokens are real billable usage.
This is important because a user may see one answer but the system may have performed much more internal work to produce it.
The economic question is therefore not whether Fugu produces a better response.
It is whether the extra quality is worth the extra computation.
For coding, research, cybersecurity and high-value enterprise workflows, that can make sense.
For a simple question, using several frontier agents would be wasteful.
Fugu’s ability to decide when not to orchestrate may ultimately be as important as its ability to orchestrate.
Pricing may become a competitive advantage or weakness
Sakana offers API subscriptions beginning at $20 per month, with higher $100 and $200 plans for heavier usage.
It also offers pay-as-you-go access.
The model creates an unusual economic problem.
If Fugu calls several expensive models internally, Sakana has to manage both quality and gross margin.
The company’s pricing rules attempt to avoid simply stacking every participating model fee on the customer.
For multi-agent Fugu requests, customers are generally charged according to the highest model tier involved rather than every model fee being added separately, while actual orchestration usage still counts.
That makes the product easier to understand.
But it means Sakana must become extremely good at routing efficiently.
An orchestrator that achieves frontier performance by using five times the compute may win benchmarks and lose the business model.
Fugu is now more than an API experiment
In August 2026, Sakana added Fugu to Sakana Chat.
That moved the product beyond developers and early enterprise users.
The updated chat product also added code execution and artifact preview capabilities.
This is significant because Fugu can now be experienced as an end-user system rather than only an infrastructure layer.
Still, the distinction between Fugu and a conventional tool-using agent matters.
Fugu’s defining capability is coordinating models.
Sakana Chat can provide tools such as code execution around it.
Those are separate layers.
Saying Fugu simply “manages tools” misses the core innovation.
It manages intelligence.
NVIDIA is now part of the strategy
In July, Sakana announced a collaboration with NVIDIA to bring NVIDIA’s open Nemotron model stack into Fugu.
NVIDIA separately confirmed the integration.
The partnership is strategically important for both companies.
For Sakana, it expands the pool of models Fugu can orchestrate and strengthens the open-model side of the platform.
For NVIDIA, it demonstrates that open models running on its accelerated-computing ecosystem can become components inside multi-agent systems rather than competing only as standalone models.
The partnership also reinforces Sakana’s positioning as a serious Japanese AI company with global infrastructure relationships.
But it does not make Fugu independent of NVIDIA.
Like most frontier AI systems, it ultimately runs on physical compute somewhere.
Model orchestration can reduce model-provider dependency.
It cannot eliminate chip and cloud dependency.
Fugu Cyber shows where orchestration may have the highest value
Sakana launched Fugu Cyber in July for cybersecurity work.
The company reports an 86.9% success rate on CyberGym and 72.1% on CTI-REALM.
Cybersecurity is a logical use case for multi-agent orchestration.
A security assessment involves different activities:
understanding a codebase, finding suspicious patterns, testing hypotheses, checking authentication, writing detection rules, verifying exploitability, producing a report.
One model does not have to be best at all of them.
A coordinator can assign specialists.
Again, the published results are Sakana’s own and should be treated as vendor-reported until independently reproduced.
But the product direction is important.
Fugu is moving from general orchestration toward vertical orchestration.
That may be where the business becomes strongest.
The real competition is not GPT versus Fugu
Fugu should not be understood as a direct replacement for GPT, Claude or Gemini.
It often uses models like those as ingredients.
That changes the competitive map.
OpenAI, Anthropic and Google are model providers.
Fugu wants to become a model-selection and coordination layer above them.
That resembles what cloud software did to infrastructure.
A database abstraction layer does not eliminate databases.
It makes it easier to use several of them.
The long-term risk for frontier-model companies is that customers stop caring which model answered the question.
If the orchestrator chooses automatically, the model becomes a replaceable backend.
That would shift bargaining power upward toward whoever owns the orchestration layer and customer relationship.
The frontier labs understand this risk, which is why they are building their own routers, agents and tool ecosystems.
Fugu’s biggest strength is also its biggest weakness
Fugu gains power from external models.
That also creates dependency.
If underlying providers raise prices, Fugu’s economics change.
If they restrict access, the pool shrinks.
If terms of service prevent certain forms of orchestration, product design may have to change.
If the best frontier models become extremely similar, there may be less value in combining them.
If one provider builds a model that dominates every important task, the need for an orchestrator could fall.
Sakana’s answer is to maintain a diverse and increasingly open model pool.
Its NVIDIA Nemotron partnership helps.
Its own Japanese models help.
But the company must prove that orchestration creates durable value rather than temporary value during a period when frontier models have unusually different strengths.
Latency is another hard problem
Multiple agents can improve quality.
They can also make users wait.
A single-model response may begin in seconds.
A multi-agent system may plan, delegate, run several model calls, verify results and synthesize a final answer.
That increases latency.
For research or code review, waiting longer may be acceptable.
For search, customer service or interactive chat, it may not be.
Fugu therefore needs accurate task triage.
The system has to know when the expensive collaborative route is justified.
This is not a minor optimization.
It is fundamental to product-market fit.
Reliability becomes harder in a multi-model system
Multi-agent systems can catch mistakes because one agent can verify another.
They can also create new failure modes.
An incorrect planner can send every worker in the wrong direction.
Agents can reinforce one another’s assumptions.
A verifier can confidently approve a bad result.
Different models can disagree in ways the orchestrator cannot resolve.
The final synthesis can omit the best evidence.
More agents do not automatically mean more truth.
The value comes from coordination quality.
That is why Sakana’s research focus on learned communication is more important than simply increasing the number of agents.
Japan’s real AI strategy is broader than Fugu
It would be inaccurate to say Fugu alone means “Japan finally entered the AI race.”
Japan’s AI ecosystem is already broader.
NVIDIA’s July announcement highlighted projects from the Institute of Science Tokyo, SoftBank, SB Intuitions, Stockmark, Hitachi, NTT DATA, ENEOS and other Japanese organizations.
Japan is building Japanese-language models, enterprise agents, robotics systems and domestic AI infrastructure.
Sakana represents one especially visible frontier-research path inside that larger effort.
Its contribution is a different architectural thesis.
Japan does not have to reproduce Silicon Valley exactly to matter.
It can compete through specialization, robotics, enterprise deployment, efficient models and orchestration.
Why Japanese enterprises could be a natural market
Japan has a large installed base of complex industrial and financial organizations.
Banks, manufacturers, trading houses, insurers and government agencies operate workflows involving large amounts of proprietary information and specialized expertise.
These organizations often cannot simply send every task to one foreign cloud model.
Compliance matters.
Data residency matters.
Auditability matters.
Vendor continuity matters.
Fugu’s ability to exclude specific agents from its pool for privacy or compliance reasons is therefore strategically relevant.
The orchestration layer could allow enterprises to use different approved models for different classes of information while maintaining one developer interface.
That is a more compelling Japanese enterprise story than building another consumer chatbot.
The future could be a market of models, not one winner
The most interesting implication of Fugu is that the AI industry may not converge on one dominant model.
Different models may remain best at different tasks.
One may dominate coding.
Another mathematics.
Another Japanese language.
Another vision.
Another cybersecurity.
Another low-cost inference.
If that happens, orchestration becomes increasingly valuable.
The winning AI product could resemble a fund manager more than a single stock.
Its advantage would come from selecting the right underlying assets and combining them intelligently.
That is Sakana’s bet.
What would prove Fugu is genuinely important
Benchmarks are only the first step.
The real evidence will come from production.
Watch enterprise customer retention.
Watch cost per solved task.
Watch latency.
Watch how often Fugu uses multiple agents versus one.
Watch whether independent benchmarks reproduce Sakana’s reported gains.
Watch whether customers obtain better results after underlying models change.
Watch how efficiently Sakana adds open models.
Watch gross margins.
Watch whether Fugu becomes infrastructure inside banks, manufacturers, research labs and government agencies.
And watch whether developers begin treating model identity as irrelevant because the orchestrator handles the choice.
That would be the strongest sign that Sakana has created a new layer of the AI stack.
Bottom line
Fugu is important because it asks a different question from most of the frontier AI industry.
Instead of asking, “How do we build the single best model?” Sakana asks, “How do we make the best available models work together?”
The answer is a language model trained to coordinate other models.
Fugu can decide when to delegate, select agents, structure collaboration, verify intermediate work and synthesize results while presenting itself to developers as one API.
Sakana’s v1.0 benchmark results are highly competitive, although they remain vendor-reported and should not be confused with independent proof of universal superiority.
Fugu Ultra v1.1 has already demonstrated the architectural advantage of the approach: when better frontier models become available, the orchestration layer can absorb them and improve without rebuilding the entire product from scratch.
The business is also becoming real.
Sakana AI is a roughly $2.65 billion startup after its 2025 Series B, has hundreds of early Fugu testers, offers production APIs and subscriptions, has brought Fugu into its consumer chat product, and is integrating NVIDIA Nemotron into the model pool.
The company still faces difficult economics.
Multi-agent inference costs money.
Latency rises.
Reliability is harder.
Dependence on external model providers does not disappear.
And benchmark leadership can vanish quickly.
But Fugu gives Japan something more valuable than another AI chatbot.
It gives the country a distinct architectural position in the global race.
The United States is spending enormous amounts to build the biggest frontier models and compute clusters.
China is pushing model efficiency, open weights and domestic hardware.
Sakana’s Japanese thesis is that the next frontier may be coordination itself.
If that thesis proves correct, the winner of the AI race may not be the company that owns the smartest individual model.
It may be the system that knows which intelligence to use, when to use it, and how to make many models behave like one.
Reader questions
Frequently asked questions
What is Fugu AI?
Sakana Fugu is a multi-agent orchestration system created by Tokyo-based Sakana AI. It behaves like one model API but can dynamically coordinate multiple underlying AI agents to solve complex tasks.
Who created Fugu AI?
Fugu was developed by Sakana AI, a Tokyo AI research company founded in 2023 by David Ha, Llion Jones and Ren Ito.
Is Fugu a Japanese AI model?
Fugu is developed by a Japan-based company, but it is best described as an orchestration model rather than a conventional standalone Japanese-language LLM. It can coordinate models from a broader agent pool.
How does Sakana Fugu work?
Fugu analyzes the request and can solve it directly or create an adaptive multi-agent workflow. It handles model selection, delegation, communication, verification and final synthesis behind a single API.
Is Fugu better than ChatGPT, Claude or Gemini?
Sakana reports highly competitive results on several benchmarks, but the Fugu scores are vendor-reported and do not prove that it is universally better. Fugu also differs structurally because it can use multiple models rather than competing only as one underlying model.
What is Fugu Ultra?
Fugu Ultra is the higher-quality variant designed for difficult multi-step tasks where answer quality is prioritized over latency and compute cost.
How much does Fugu cost?
Sakana currently lists Fugu Ultra at $5 per million standard input tokens and $30 per million output tokens, with orchestration tokens also counted as billable usage. Subscription plans are also available.
What is Fugu Cyber?
Fugu Cyber is a specialized orchestration model for cybersecurity. Sakana reports success rates of 86.9% on CyberGym and 72.1% on CTI-REALM, although those are company-reported benchmark results.
Why is Fugu important for Japan?
Fugu gives Japan a distinctive frontier-AI strategy centered on model orchestration, specialization and reduced single-vendor dependency rather than simply trying to match the largest U.S. and Chinese companies in brute-force model scale.
Is NVIDIA involved with Sakana Fugu?
Yes. Sakana AI and NVIDIA announced that NVIDIA Nemotron open models are being integrated into Fugu’s model pool.
Nexuswild welcomes factual corrections. Email [email protected] with evidence and the article URL.
