TLDR
- Sakana AI says Fugu Ultra matches Claude Fable 5 and Mythos on difficult benchmarks today.
- Fugu uses one API endpoint to coordinate specialized models across coding, research, and reasoning tasks.
- The system can delegate work, verify outputs, and merge responses into one user-facing answer.
- Fugu Ultra targets complex workloads, including AI research, cybersecurity, patent search, and financial forecasting.
- Sakana AI plans subscriptions and pay-as-you-go pricing for consumer, enterprise, and high-volume users globally
Japanese artificial intelligence startup Sakana AI has introduced Fugu, a multi-agent system designed to coordinate other AI models through a single interface. The Tokyo-based lab said the system is built to decide whether a task should be handled directly or assigned across multiple specialized models. The company presents Fugu as a unified API that can manage routing, delegation, review, and response assembly.
Sakana AI was co-founded by researchers with backgrounds in major AI research work, including one contributor to the original Transformer paper. The Transformer architecture remains central to many modern large language models used in coding, search, reasoning, and text generation. The company’s latest release places its focus on orchestration rather than only increasing the size of one model.
The product line includes Fugu and Fugu Ultra, according to the information released by the company. Fugu is positioned for faster coding, chat, and lower-latency everyday use. Fugu Ultra is aimed at harder tasks requiring multi-step reasoning, technical analysis, and coordination across expert systems.
Fugu Ultra benchmark claims draw attention
Sakana AI says Fugu Ultra matches Claude Fable 5 and Mythos Preview on difficult engineering, science, and reasoning benchmarks. The company also claims the system can outperform Gemini 3.1 Pro, Opus 4.8, and GPT-5.5 on selected tasks such as AutoResearch, mechanical design, and financial forecasting. These performance statements are presented as company claims and were not independently verified here.
The system differs from a conventional single-model strategy because it can organize a pool of models behind one request. When a user sends a prompt, Fugu may answer alone or assign different parts of the problem to specialized systems. It can then review the results and combine them into one response for the user.
The company describes the approach as useful for complex workflows where one model may not be the strongest option for every step. Tasks such as research reproduction, patent review, cybersecurity analysis, and technical design often require several forms of reasoning. Fugu Ultra is being marketed toward those heavier workloads rather than only general chat use.
Unified API targets enterprise workloads
Both Fugu and Fugu Ultra are planned for access through a unified API, according to Sakana AI’s release details. The setup is intended to let users interact with one endpoint while the system manages model selection in the background. That structure could appeal to developers and companies seeking simpler access to multi-model workflows.
The company also said the model pool is designed to be swappable when outside access changes. This means Fugu may route tasks through available providers if access to a specific model becomes restricted. The design is being presented as a way to reduce dependency on any one external model provider.
Sakana AI plans subscription options for everyday users and pay-as-you-go pricing for high-volume and enterprise workloads. The pricing structure is aimed at supporting both individual access and heavier commercial usage. The launch places the Japanese AI lab in the competitive market for advanced reasoning systems, where model coordination is becoming a separate product strategy.
Get 3 Free Stock Ebooks
Discover top-performing stocks in AI, Crypto, and Technology with expert analysis.
- Top 10 AI Stocks - Leading AI companies
- Top 10 Crypto Stocks - Blockchain leaders
- Top 10 Tech Stocks - Tech giants
