Sakana AI's Fugu Bets on Coordinating Leading Models to Challenge AI Giants

A Tokyo Startup Puts AI Coordination to the Test

TOKYO, July 17, 2026. Sakana AI is challenging the race to build ever larger artificial intelligence models with Fugu, a service that coordinates several leading systems behind a single interface. Unveiled on June 22, the platform uses a central coordinator estimated at roughly 7 billion parameters to decide which models should tackle a task, check their work and combine their answers.

Sakana AI says its premium service, Fugu Ultra, scored 73.7 on SWE-Bench Pro, a test of fixing bugs in real software repositories, compared with 69.2 for Claude 4.8 Opus and 58.6 for GPT-5.5. Independent verification remains limited, but the company is using those results and feedback from nearly 500 beta testers to make its case for a different approach to advanced AI.

Fugu, named after the Japanese word for pufferfish, appears to developers as one model. Behind that interface, it can organize a team of specialized AI systems, including models from OpenAI, Anthropic and Google. Its central proposition is simple: a smaller system might achieve powerful results by knowing which other systems to ask, rather than storing every capability within itself.

The launch also addresses a business concern beyond answer quality. Earlier in June, Anthropic abruptly revoked public access to Claude Mythos 5 and Claude Fable 5 following a United States government export control order. International developers who had built workflows around those models faced a sudden loss of access. Sakana AI explicitly cites such disruptions as a reason to build a service that can change its underlying providers.

Fugu therefore carries two connected ambitions: to compete with individual frontier models on demanding tasks and to give customers a way to spread their reliance across providers. Its commercial appeal depends on how well that coordination works, how much it costs and which models remain available.

Why Sakana AI Is Taking a Different Route

OpenAI, Anthropic, Google and other leading developers have largely pursued progress through larger foundation models, more training data and greater computing power. A foundation model is a general purpose AI system that can support many applications. Parameters are the numerical settings learned during training that shape how a model responds.

That approach has produced systems capable of passing bar examinations, writing sophisticated code and performing extended reasoning. It has also concentrated access to the strongest models among a handful of American providers. Export controls, company policy changes and geopolitical disputes can alter access quickly, creating exposure for businesses elsewhere.

Sakana AI, founded in 2023 by former Google Brain researcher David Ha and Llion Jones, is pursuing collective intelligence instead. Jones co-authored the 2017 transformer paper Attention Is All You Need, which helped establish the architecture behind modern language models. Public disclosures indicate that Japanese megabanks and chipmaker Nvidia backed a funding round valued at roughly 30 billion yen.

Those resources give the Tokyo startup room to compete with much larger businesses without trying to reproduce their entire training strategy. Rather than building a model with a trillion parameters, Sakana AI has designed a smaller coordinator to assemble capabilities already available elsewhere.

The developments leading to the current launch span several dates:

• 2017: Llion Jones co-authored Attention Is All You Need.
• 2023: David Ha and Jones founded Sakana AI.
• 2026: The Trinity and Conductor research papers were presented at the International Conference on Learning Representations.
• Earlier in June 2026: Anthropic revoked public access to Claude Mythos 5 and Claude Fable 5 after a United States export control order.
• June 22, 2026: Sakana AI unveiled Fugu.
• End of July 2026: The deadline for joining the subscription promotion that provides a free second month.

What Happens Behind a Single Request

A developer sends Fugu a request through an application programming interface, or API, compatible with OpenAI standards. An API is the connection software uses to request work from another service. That familiar interface allows Fugu to present a coordinated workflow as if it were a call to one language model.

Sakana AI describes the central model as a reinforcement learning conductor. It examines the request and decides whether to answer directly or distribute pieces of the task among specialist agents. In most cases, the conductor organizes the work instead of producing the final answer by itself.

The interchangeable agent pool includes frontier models from OpenAI, Anthropic and Google, as well as instances of Fugu itself. Frontier models are the leading systems at the edge of current capability. An agent, in this setting, is a model assigned a particular job within a larger workflow.

The coordinator operates like a foreman: it considers what needs doing, assigns suitable workers, checks their outputs and assembles the result. A complex request can therefore involve several systems even though the developer receives one unified response.

Fugu can also send smaller tasks back to itself. This recursive capability creates several layers of reasoning when a problem needs extended analysis. The design is intended to manage workflows of arbitrary complexity without requiring the coordinator to contain the encyclopedic knowledge held in frontier models with hundreds of billions of parameters.

Sakana AI's argument is that organizational intelligence can compensate for a smaller model. The coordinator needs to know how to use the capabilities around it, while the specialist systems supply much of the knowledge and execution.

The Research Behind the Conductor

Two papers presented at the 2026 International Conference on Learning Representations provide the research foundation: 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 coordinator creates and changes agent structures as a task develops. Evolutionary computing explores candidate solutions and modifies them to find arrangements that work better. Here, that process helps determine how different AI agents should be organized.

Trinity assigns roles such as Thinker, Worker and Verifier. Thinkers analyze what the task requires, Workers perform specific computations and Verifiers examine outputs for accuracy and consistency. Those roles separate planning, execution and checking within the same workflow.

Conductor builds on that approach through deep reinforcement learning. This is a training method in which a system learns strategies through feedback on its actions. The coordinator learns when to delegate, when to answer directly, how agents should communicate and how their outputs should be merged.

The resulting workflow is adaptive rather than a fixed sequence written in advance. Sakana AI says the system discovers coordination strategies through experience, giving it a way to address unfamiliar task types that its designers may not have anticipated. The research aims to make delegation itself a learned capability.

What the Benchmark Scores Show

Sakana AI's published evaluations place Fugu Ultra ahead of the comparison models on SWE-Bench Pro. Its score of 73.7 exceeds Claude 4.8 Opus at 69.2 by 4.5 points and GPT-5.5 at 58.6 by 15.1 points. The test asks AI systems to find and repair bugs in actual software repositories, making it relevant to one of the service's main proposed uses.

The company also gives scores across questions, coding, computer operations and difficult reasoning:

• GPQA-Diamond: 95.5 on a test of questions at graduate level designed to resist simple web searches.
• LiveCodeBench: 93.2 on a coding benchmark.
• TerminalBench 2.1: 82.1 on a benchmark evaluating system administration capabilities.
• Humanity's Last Exam: 50.0 on a difficult reasoning test.

Independent verification of these figures remains limited. Testing conditions, the techniques used to write prompts and the number of attempts allowed can affect scores. The AI industry has also seen benchmark results that are difficult to reproduce in production, where tasks and operating conditions vary.

The range of company results suggests that orchestration can compete with individual frontier models across several disciplines, with stronger performance in some areas. Sustained external scrutiny and performance in customers' actual workflows will determine how far that advantage extends.

For buyers, the important comparison combines answer quality with the time and computing resources required to produce it. A system that asks several models to work on a problem may find more issues, while also taking longer and consuming more resources than a simpler request.

What Nearly 500 Beta Testers Reported

Nearly 500 beta testers used Fugu during an extended preview focused on long, complex workflows. Software development feedback centered on the thoroughness of Fugu Ultra's code reviews, particularly its ability to identify problems missed by baseline models.

One engineer participating in the beta told Sakana AI that baseline models typically flagged around three issues in a given codebase, while Fugu found more than twenty. The engineer said it also supplied detailed explanations and steps to address the problems.

The beta engineer described Fugu Ultra as the preferred system for reviewing code:

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.

Cybersecurity teams also used Fugu Ultra to automate complete penetration testing cycles. Penetration testing is an authorized attempt to find weaknesses in a system. A human operator defined the scope, and the service carried out reconnaissance, tested vulnerabilities, reviewed authentication mechanisms and prepared a report.

The security work included checks for XSS and SQL injection. XSS involves making a website run unwanted scripts, while SQL injection involves manipulating database commands through an application's inputs. Fugu's final reports included evidence and exact procedures for testing again after repairs.

Beta participants said the system stayed within its operational limits and avoided destructive actions while maintaining thoroughness. Enterprise executives also highlighted persona stability: Fugu maintained a consistent identity and focus during extended sessions in which other models often suffered context degradation, or a weakening grasp of earlier instructions and information.

Two Tiers With Different Controls

Sakana AI is selling standard Fugu and Fugu Ultra through one API endpoint compatible with OpenAI standards. The two tiers target different balances of speed, quality and customer control.

Standard Fugu is optimized for low latency production tasks. Latency is the delay between making a request and receiving a response. Its intended uses include interactive chatbots, live coding assistance and code review tools such as Codex, where users expect nearly instantaneous responses.

The standard tier allows customers to exclude specific underlying agents from the routing pool manually. For organizations subject to strict privacy or data governance rules, that control provides a way to prevent sensitive information from reaching particular providers while continuing to use the orchestration service.

Fugu Ultra is designed to prioritize answer quality on demanding tasks involving many reasoning steps. Sakana AI lists academic paper reproduction, patent investigations, advanced data science workflows and exhaustive security audits among its intended applications.

Ultra uses a deeper, fixed pool of expert agents, so customers cannot exclude individual agents as they can with the standard service. Its current model identifier is fugu-ultra-20260615. Greater coordination means higher latency and more token consumption, a trade Sakana AI argues is worthwhile for applications where accuracy is critical.

The difference matters for regulated businesses. Standard Fugu offers control over which agents can receive a task; Ultra offers its fixed configuration for deeper work. Customers must weigh those choices against their rules for handling information and their tolerance for slower responses.

Pricing, Usage and the European Gap

Fugu Ultra costs $5 per million input tokens and $30 per million output tokens. Tokens are the units of text processed by a model, such as words or parts of words. Input tokens cover material sent to the system, while output tokens cover its response.

The rates double for contexts longer than 272,000 tokens. A context is the information available to the system while it works, including the request and supporting material. At those longer lengths, the listed rates become $10 per million input tokens and $60 per million output tokens.

Subscription plans range from $20 to $200 per month. Customers who subscribe before the end of July 2026 receive a free second month at their selected tier. The service reports token usage and associated costs for each request, allowing engineering and finance teams to monitor spending in real time.

Fugu is available globally except in the European Union and European Economic Area while Sakana AI works to align its operations with the General Data Protection Regulation, or GDPR. No date has been given for availability in those markets. The restriction reduces the platform's international reach at launch.

Cost control is especially relevant for Ultra's intended workloads. Long investigations and repeated checks can use substantial computing quotas. The price of the response is therefore part of the decision about whether a deeper review is worthwhile for a particular task.

Resilience Against Losing a Model Provider

Anthropic's June withdrawal of public access to Claude Fable 5 and Claude Mythos 5 made provider dependence a practical concern for international customers. Workflows built around those systems lost access overnight, leaving procurement teams without an immediate replacement.

Sakana AI presents Fugu as insurance against that kind of disruption. Its routing architecture can shift work away from unavailable or restricted providers and toward other models. If Anthropic systems become inaccessible, the company says workloads can move to OpenAI, Google or openly available alternatives.

A new model that performs especially well on a task can also be added to the pool without requiring customers to rewrite applications around another API. The customer keeps the same interface while Sakana AI changes the systems doing the work underneath.

The company connects that flexibility to AI sovereignty, meaning greater control over access to AI capabilities. For Japanese enterprises and government institutions, domestic capability has become a strategic priority amid economic and national security concerns about relying entirely on American frontier systems.

A Tokyo vendor offering local support, regional data handling options and a different technical approach appeals to buyers seeking to diversify their technology suppliers. Sakana AI's backing from major Japanese financial institutions strengthens that position.

Fugu still depends heavily on the American providers it is designed to route around. If access to GPT, Claude and Gemini were all severed, it would lack the systems supporting its claimed frontier results. Fuller independence would require domestic or openly available models capable of matching them, and that ecosystem does not yet exist at scale.

Critics also point out that Fugu is a proprietary commercial service rather than an open framework. Customers move their relationship to Sakana AI, accepting reliance on another vendor with a different business model. The practical benefit is diversification beneath one interface, alongside a new service provider responsible for keeping that interface working.

What Fugu Could Change About AI Development

AI development has spent much of the past decade following a scaling hypothesis: more parameters, more data and more computing power produce better results. The strategy has delivered impressive capabilities, while raising entry costs to levels that only the wealthiest companies and governments can readily afford.

Training a frontier foundation model costs hundreds of millions of dollars and requires specialized hardware that is itself subject to export restrictions. Sakana AI's approach shifts attention toward how existing models are organized and how efficiently they are used.

A relatively small coordinator trained specifically to delegate could give organizations with smaller budgets a way to build powerful systems without training a GPT-class foundation model from scratch. The competitive question becomes whether architectural design and coordination can deliver enough additional value to justify their operational costs.

Researchers have long explored multiagent systems and mixtures of experts, approaches that distribute work among specialized components. Fugu is among the first prominent commercial implementations presented as a direct alternative to individual frontier models. Its performance could influence whether investment flows toward coordination technologies alongside foundation model scaling.

The beta feedback points particularly toward code review and security auditing, where tasks stretch across many steps and thorough checking matters. These workflows may benefit from distributed work more than applications that need a fast answer from a single model call.

Fugu's limitations remain part of that calculation. Complex coordination introduces delays that can make it unsuitable for applications requiring responses within milliseconds. Intensive tasks consume substantial computing quotas, and opaque routing can make it difficult for customers to know which underlying model handled their data.

Sakana AI acknowledges those constraints while arguing that the benefits justify them across analytical and engineering workloads. As of July 17, its case rests on company benchmark results, beta feedback and the value of switching among providers. The central test is whether that combination delivers reliable quality at a cost and speed customers can accept.

Key Points

• Sakana AI unveiled Fugu on June 22, 2026, using a coordinator estimated at roughly 7 billion parameters to organize other AI models.
• The service draws on the Trinity and Conductor research papers presented at ICLR 2026 and supports recursive delegation.
• Sakana AI gives Fugu Ultra a SWE-Bench Pro score of 73.7, compared with 69.2 for Claude 4.8 Opus and 58.6 for GPT-5.5; independent verification remains limited.
• Nearly 500 beta testers tried the service, with feedback highlighting code review, security testing and stability during extended sessions.
• Standard Fugu prioritizes speed and allows agent exclusions. Ultra uses a fixed, deeper pool for demanding reasoning tasks.
• Ultra costs $5 per million input tokens and $30 per million output tokens, with rates doubling above 272,000 context tokens.
• Subscriptions range from $20 to $200 monthly, with a free second month for customers joining before the end of July 2026.
• Fugu is unavailable in the EU and EEA pending GDPR alignment and continues to depend heavily on American frontier model providers.