A New Playbook for Artificial Intelligence
TOKYO — On June 22, 2026, Sakana AI unveiled Fugu, a multiagent orchestration service that poses a provocative question to the technology industry. Does the next leap in artificial intelligence require building an even larger monolithic model, or can a smaller, smarter system achieve similar results simply by knowing whom to ask? Fugu, named after the Japanese word for pufferfish, represents the Tokyo startup bet on the latter approach. The platform behaves like a single model to the outside world while internally coordinating a team of specialized frontier artificial intelligence systems to solve complex tasks.
The announcement arrives at a critical inflection point for enterprise adoption of generative AI. For years, progress has been measured largely by parameter counts, training compute, and the race to build bigger foundation models. OpenAI, Anthropic, Google, and others have pursued this scaling philosophy with remarkable results, producing systems capable of passing bar exams, writing sophisticated code, and conducting extended reasoning. Yet this concentration of capability in a handful of American providers has created new vulnerabilities. Access to the most powerful models can change overnight because of export controls, corporate policy shifts, or geopolitical disputes, leaving international businesses exposed.
Sakana AI proposes an alternative architecture rooted in collective intelligence. Rather than attempting to match frontier labs by training a trillion-parameter behemoth, the company has built a compact coordinator, estimated at roughly 7 billion parameters, that functions as a master delegator. When a user submits a request through the Fugu API, which is compatible with OpenAI standards, the system analyzes the task and either answers directly or assembles a team of expert models to handle specific sub-tasks. It then verifies their outputs and synthesizes a unified response. To the developer, the experience resembles calling a single large language model. Underneath, an entire committee of agents may have contributed to the answer.
The timing could hardly be more strategic. Earlier in June, Anthropic abruptly revoked public access to the Claude Mythos 5 and Claude Fable 5 models following a United States government export control order. The move sent ripples through the developer community, particularly among international clients who had built workflows around those systems. Sakana AI explicitly cites such disruptions as motivation for the Fugu design, positioning the service as a hedge against dependency on a single vendor and a step toward what it calls AI sovereignty.
Founded in 2023 by former Google Brain researcher David Ha and Llion Jones, who co-authored the seminal 2017 transformer paper Attention Is All You Need, Sakana AI has attracted significant attention and funding. Public disclosures indicate that Japanese megabanks and chipmaker Nvidia backed a funding round valued at roughly 30 billion yen, giving the startup resources to compete against far larger rivals. With Fugu, the company aims to prove that Japanese innovation can challenge the frontier model paradigm not by copying it, but by reimagining how existing capabilities are assembled.
Inside the Orchestration Engine
Understanding Fugu requires looking past the surface of its simple API. When a prompt arrives, the system passes it to a central model that Sakana AI describes as a reinforcement learning conductor. This conductor does not generate the final answer itself in most cases. Instead, it plays the role of a foreman on a construction site, dynamically analyzing the task type and assigning pieces of work to the most suitable agents available in an interchangeable pool. That pool includes third-party frontier models from providers like OpenAI, Anthropic, and Google, as well as recursive instances of Fugu itself for nested reasoning.
The theoretical foundation for this behavior comes from two research papers presented at the 2026 International Conference on Learning Representations: Trinity: An Evolved LLM Coordinator and Learning to Orchestrate Agents in Natural Language with the Conductor. Trinity introduces an evolutionary computing framework in which the central coordinator creates and modifies agent structures on the fly, labeling them with roles such as Thinker, Worker, or Verifier depending on the demands of the task. Thinkers analyze requirements, Workers execute specific computations, and Verifiers check outputs for consistency and accuracy.
Conductor builds on this by employing deep reinforcement learning to train the coordinator in the art of delegation in natural language. Through extensive training, the small central model learns when to delegate versus when to answer directly, how agents should communicate with one another, and how to merge disparate outputs into a coherent final product. The result is not a predetermined workflow but an adaptive system that discovers coordination strategies through experience. This distinction matters because it allows Fugu to handle novel task types that its designers may not have explicitly anticipated, rather than being limited to rigid playbooks.
From an engineering perspective, one of the most intriguing aspects of Fugu is its recursive capability. The orchestrator can dispatch sub-tasks back to itself, creating multilayered reasoning chains for problems requiring extended analysis. This means the 7 billion parameter coordinator can manage workflows of arbitrary complexity without itself possessing the encyclopedic knowledge stored in the hundreds of billions of parameters found in frontier foundation models. It compensates for its smaller size with superior organizational intelligence, a concept that challenges the prevailing assumption that bigger is always better.
Benchmarks and Beta Testing
Sakana AI has accompanied the Fugu launch with a bold set of performance claims. On SWE-Bench Pro, a widely respected benchmark that tests the ability of an AI system to locate and fix bugs in actual software repositories, Fugu Ultra reportedly scored 73.7. According to published evaluations from the company, that places it ahead of Claude 4.8 Opus at 69.2 and GPT-5.5 at 58.6. Additional reported scores include 95.5 on GPQA-Diamond, a graduate-level question-answering test designed to be resistant to simple web search; 93.2 on LiveCodeBench; 82.1 on TerminalBench 2.1, which evaluates system administration capabilities; and 50.0 on Humanity’s Last Exam, a notoriously difficult reasoning test.
Independent verification of these figures remains limited, and the AI industry has a long history of benchmark results that prove difficult to reproduce in production environments. Testing conditions, prompt engineering techniques, and the number of attempts allowed can all influence scores. Nevertheless, the breadth of disciplines covered in reports issued by the company suggests that the orchestration approach is at least competitive with monolithic frontier models, if not superior in certain domains.
More revealing than the benchmarks may be the feedback from nearly 500 beta testers who used Fugu during an extended preview period focused on long, complex workflows. In software development, participants reported that Fugu Ultra demonstrates exceptional thoroughness during code reviews. One engineer participating in the beta told Sakana AI that while baseline models typically flagged around three issues in a given codebase, Fugu surfaced more than twenty, providing detailed explanations and remediation steps.
For code review, Fugu Ultra is better than GPT-5.5. It gives comprehensive answers and finds the bugs others miss. It has become the model I run all my reviews through.
Beyond coding, cybersecurity teams deployed Fugu Ultra to automate complete penetration testing cycles. A human operator would define the scope, and the orchestration engine would execute reconnaissance, test for XSS and SQL injection vulnerabilities, review authentication mechanisms, and compile a final report with evidence and exact retest procedures. Beta participants confirmed the system adhered strictly to its operational parameters, avoiding destructive actions while maintaining thoroughness. Enterprise executives also highlighted persona stability, noting that Fugu maintained consistent identity and focus across extended sessions where other models often suffered from context degradation.
Two Tiers, One API
Sakana AI is commercializing Fugu through two distinct service tiers, both accessible via a single API endpoint compatible with OpenAI standards. The standard Fugu model is optimized for low latency production tasks such as interactive chatbots, live coding assistance, and code review tools like Codex. It balances performance with speed, making it suitable for applications where users expect nearly instantaneous responses. For organizations operating under strict data governance or privacy regulations, the standard tier allows specific underlying agents to be manually opted out of the routing pool. This feature enables compliance teams to prevent sensitive data from reaching particular third-party providers without abandoning the orchestration framework entirely.
Fugu Ultra occupies the premium tier. It is engineered for maximum answer quality on complex, demanding tasks that require deep reasoning across many steps. These include academic paper reproduction, patent investigations, advanced data science workflows, and exhaustive security audits. Ultra coordinates a deeper and fixed pool of expert agents, which means the opt-out functionality is not available. The trade-off is higher latency and greater token consumption, but Sakana AI argues that the gains in accuracy justify the cost for mission critical applications. The current model identifier for this tier is fugu-ultra-20260615.
Pricing reflects the distinction between the two products. Fugu Ultra costs $5 per million input tokens and $30 per million output tokens, with rates doubling for contexts longer than 272,000 tokens. Subscription plans range from $20 to $200 per month, and Sakana AI is offering subscribers who join before the end of July 2026 a free second month at their chosen tier. Token usage and associated costs are reported per request, allowing finance and engineering teams to monitor spending in real time. Notably, the service is not currently available in the European Union or European Economic Area while the company works to align its operations with the General Data Protection Regulation, a gap that limits the global reach of the platform at launch.
The Geopolitics of Model Routing
The commercial pitch behind Fugu extends beyond technical architecture into the arena of geopolitical risk management. The modern AI supply chain is highly concentrated. A small number of American companies control access to the most capable foundation models, and their decisions are increasingly subject to government oversight. When the United States imposed export controls on Anthropic Fable and Mythos models in June 2026, the episode served as a warning for enterprises outside America that had built critical workflows around a single provider. Overnight, access disappeared, and procurement teams had no immediate alternative.
Sakana AI is pitching Fugu as an insurance policy against exactly this scenario. Because the orchestrator can dynamically route around unavailable or restricted providers, the company argues that customers gain resilience without sacrificing capability. If Anthropic models become inaccessible, Fugu shifts the workload to OpenAI, Google, or openly available alternatives. If a new model enters the market and excels at a particular task, Sakana AI says it can be folded into the pool without forcing customers to rewrite their applications around a new API. This interchangeable architecture represents a more concrete form of AI sovereignty than many of the abstract policy discussions currently circulating in technology circles.
For Japanese enterprises and government institutions, the pitch carries particular resonance. Japan has made domestic AI capability a strategic priority, and trusting entirely on American frontier stacks creates both economic and national security concerns. A vendor based in Tokyo offering local support, regional data handling options, and an alternative technical philosophy provides a compelling narrative for buyers seeking to diversify their technology portfolios. The investor base of Sakana AI, which includes major Japanese financial institutions, reinforces this positioning.
Yet skepticism is warranted. The Fugu agent pool remains heavily dependent on the very American giants it claims to circumvent. The system cannot deliver frontier-level results if API access to GPT, Claude, and Gemini is severed entirely. True independence would require a robust ecosystem of domestic or openly available models capable of matching those systems, and that ecosystem does not yet exist at scale. Critics also note that Fugu is a proprietary commercial service rather than an open framework, which means customers are trading dependence on one vendor for dependence on another, albeit one with a different business model.
What the Launch Means for AI Development
If the Fugu performance claims hold up under sustained external scrutiny, the consequences for artificial intelligence research could be significant. For the better part of a decade, the field has pursued a scaling hypothesis: more parameters, more data, and more compute yield better results. This philosophy has produced genuinely impressive systems, but it has also created barriers to entry so high that only the wealthiest corporations and governments can participate. Training a frontier foundation model costs hundreds of millions of dollars and requires specialized hardware that is itself subject to export restrictions.
Fugu points to a different path. It suggests that a relatively small model, trained specifically for coordination rather than encyclopedic knowledge, can punch above its weight by intelligently using existing resources. This shifts the competitive landscape from raw training capacity to architectural ingenuity and orchestration efficiency. It also democratizes access to high-performance AI in the sense that a skillfully designed router can theoretically be built by organizations with far smaller budgets than those required to train GPT-class models from scratch.
The research community has long explored multiagent systems and mixtures of experts, but Fugu represents one of the first prominent commercial implementations positioned as a direct alternative to monolithic frontier models. Its success or failure will likely influence whether venture capital and corporate research dollars begin flowing toward orchestration and coordination technologies rather than exclusively toward foundation model scaling. Already, the beta feedback suggests that certain enterprise workflows, particularly those involving tasks extending over long horizons like code review and security auditing, may be better served by distributed systems than by models invoked in a single call.
Of course, the approach is not without trade-offs. The complex multiagent reasoning steps can introduce latency that makes Fugu unsuitable for applications requiring millisecond responses. High-intensity tasks also consume significant compute quotas, and the opaque nature of the routing logic means users cannot always know which underlying model handled their data, complicating compliance efforts in regulated industries. Sakana AI has acknowledged these limitations while arguing that the benefits outweigh the costs for a broad swath of analytical and engineering workloads.
Key Points
- Sakana AI launched Fugu, a multiagent orchestration service that routes tasks across an interchangeable pool of frontier AI models through a single API compatible with OpenAI standards.
- The system relies on a compact coordinator model of roughly 7 billion parameters rather than a massive monolithic foundation model, using research from the Sakana AI Trinity and Conductor papers.
- Fugu Ultra claims competitive benchmark scores on SWE-Bench Pro, GPQA-Diamond, LiveCodeBench, and other tests against leading systems including GPT-5.5 and Claude Opus 4.8.
- Two tiers are available: standard Fugu for low latency production tasks with manual opt-out compliance controls, and Fugu Ultra for complex reasoning with a fixed deeper agent pool.
- The launch targets enterprises seeking resilience against vendor lock-in and export control disruptions, though the service remains dependent on American frontier model APIs.
- Fugu is available globally except in the EU and EEA, where GDPR compliance is pending; pricing for Ultra starts at $5 per million input tokens and $30 per million output tokens.
