/ The company announced latest AI, Fugu and Fugu Ultra.
Mike Alvarez Joined TheTweaks as an emerging tech , startup and launches journalist. He comes up with those things or products which are not there yet and are hardly possible to come close to. He covers the latest technology, startups and new product launches and releases. He worked 4 years in a startup company in Austin and later he joined TheTweaks and he is proving himself here with his commendable writing performance. He is also the publisher of a small consumer robotics newsletter.
The company announced Fugu and Fugu Ultra, the latest artificial intelligence solutions capable of solving complicated tasks related to reasoning, coding, and research using the multi agent technology. The presentation of Fugu and Fugu Ultra took place on June 22, 2026.
Sakana AI becomes one of the leading players in AI market exploring alternative approaches to designing systems.
Fugu and Fugu Ultra use the multi agent orchestration system allowing multiple AI models to cooperate to perform complicated tasks. Unlike single large language models generating responses based on a single neural network, Fugu uses coordination layer where each part of a certain problem is assigned to a separate model.
The core of the system is a 7 billion parameter orchestrator model acting as a decision making engine. The orchestrator generates no responses, selecting the external AI models from the list, delegating subtasks, evaluating intermediate outputs, and finally generating the final response.
Thus, Fugu is not a simple chatbot but a full fledged system level AI solution where multiple sources of intelligence interact.
Fugu Ultra is a more sophisticated version of the product designed to perform long horizon reasoning, complex coding tasks, and advanced analytical workloads. The product is optimized for multi step problem solving, where several models need to communicate in a sequence or simultaneously.
The system was announced on June 22, 2026, being one of the most notable updates in Sakana AI’s research into collective intelligence systems.
The main motivation of developing Fugu is the inefficiency of scaling single models. As models become larger, their training, computing power, and energy consumption increase dramatically. Thus, Sakana AI’s approach aims to distribute intelligence between several specialized systems instead of having one massive model.
Another motive is the ability to improve resiliency and independence of the system in regard to single model dependence. In this way, Fugu is able to perform tasks dynamically choosing several models and avoiding possible failures caused by model failures, API limitations, or other external constraints such as legal or export regulations.
Fugu is based on the research contributions described in two ICLR 2026 papers: TRINITY and The Conductor.
TRINITY proposes a lightweight evolved coordinator assigning different roles to AI systems. These include Thinker, Worker, and Verifier roles. Such approach helps to break down complicated tasks into components.
The Conductor suggests using reinforcement learning to discover coordination strategies in natural language. In this way, the orchestration efficiency gets improved over time.
Therefore, these two research developments make up the backbone of Fugu’s capabilities to orchestrate multiple models.
Practically, when a user requests a query, Fugu splits it into subtasks, sends to different models, evaluates their outputs, and generates the final response. Such layered architecture allows handling such complex workflows as coding, reasoning, and research analysis in some cases better than a single model.
The company released the benchmarks demonstrating the competitive positioning of Fugu Ultra among frontier AI systems.
On SWE Bench Pro coding benchmark, Fugu Ultra scores 73.7 that is higher than 69.2 score of Claude Opus 4.8. On Humanity’s Last Exam, the benchmark designed to measure advanced reasoning capabilities, Fugu Ultra scores 50.0 that is slightly higher than 49.8 score of Claude Opus 4.8.
Thus, the multi agent orchestration approach is able to compete with and sometimes to surpass the traditional approach based on large models in specialized tasks.
Sakana AI has revealed the prices of using Fugu Ultra. The solution costs $5 per million input tokens and $30 per million output tokens, with double prices for the contexts exceeding 272K tokens.
The API is OpenAI compatible and will allow for easy integration into the existing applications without any infrastructure changes.
Still, there are geographical limitations. Currently, the API is not available in the European Union or European Economic Area while Sakana AI is working on becoming compliant with GDPR regulations.
The launch of Fugu and Fugu Ultra shows how the paradigm of AI system design changes. Instead of competing in the size of models, the companies start looking for innovative approaches to improve performance and efficiency of AI systems.
Thus, using multiple specialized models instead of one large model allows Sakana AI to reduce computational waste and to make task routing more efficient. Besides, the system can become more adaptive using the ability to switch different models without retraining.
Such approach makes Fugu particularly relevant for enterprise solutions where reliability, cost control, and flexibility are vital.
There are several main advantages of the Fugu architecture.
First of all, it increases the resiliency of the system making it possible to work despite failure or unavailability of one of the models. It also enhances the flexibility of the system making it possible for different models to specialize in different tasks.
Moreover, it reduces the dependence on any single AI provider that can be very useful in constrained environments.
Finally, it introduces the new approach to scaling AI systems, which is not dependent on the growing of the models’ size.
Limitations
Nevertheless, the multi agent system also raises the level of complexity. For successful cooperation of multiple models, there should be an effective orchestration layer. Any error in routing or synthesis may negatively affect the output quality.
Besides, there are issues with the costs. The usage of multiple external models in one workflow can increase the cost of operations depending on the workload.
Also, there is an issue with the performance consistency because the output quality will depend on the availability and reliability of external models in the system’s pool.
This development by Sakana AI is one of the examples of the current trend of agent based AI systems and modular architecture. In this way, instead of scaling the transformer models, researchers try to find ways of combining multiple smaller systems into a coherent intelligence network.
This approach may influence the way the future AI platforms will be designed, especially in the enterprise environment where reliability and adaptability are as important as the performance.
According to TheTweaks, the launch of Fugu and Fugu Ultra demonstrates the evolution in the sphere of AI system design. Now, the focus is shifting from “bigger models” to “smarter systems” where coordination of multiple agents is a key point.
Recently, TheTweaks covered this topic publishing articles related to the research of AI orchestration, enterprise AI infrastructure, and modular model design.
Tim Cook’s 15 years as Apple’s chief executive officially ended on September 1 and now the torch has been passed on to John Ternus, who is the company’s long-time hardware. It’s Apple’s first leadership change…
Be respectful and constructive. Have a question or feedback? We’d love to hear from you. Contact us at contact@thetweaks.com