The world of artificial intelligence has focused primarily on the race for parameter count and training data size in recent months. Giants like Google or OpenAI are betting on ever larger models that are supposed to "know everything." However, the Japanese company Sakana AI comes with a different philosophy. Their system Fugu does not try to be the largest model in the world, but strives to be the smartest system composed of many smaller, specialized parts.

What is the Fugu system and what makes it different?

To understand the significance of Sakana AI's announcement, we must first explain how a traditional LLM (Large Language Model) works. Most current models are "monolithic" – they are one huge network that tries to solve everything from writing poems to programming in Python. This is extremely demanding on computational power and energy.

Sakana AI uses technology that experts call evolutionary model merging and multi-agent systems. The Fugu system works more like an orchestra than a single musician. Instead of one model solving the entire problem, Fugu can dynamically select several smaller, specialized models and combine their capabilities for a specific task. The result is a system that is more efficient than the largest models on the market in certain domains – for example, in logical reasoning or mathematics.

According to reports from NDTV, Fugu shows extraordinary results in benchmarks that test deep contextual understanding and the ability to solve new, unknown problems. This is precisely an area where monolithic models often "get stuck" due to their tendency to merely reproduce learned patterns.

Benchmarks: Fugu vs. GPT-5.5 and Claude 5

Comparison with market leaders is very ambitious in this case. Available data suggests that Fugu does not beat the competition in every metric (which would be unrealistic), but in specific categories such as algorithmic reasoning and inference efficiency.

• GPT-5.5 (OpenAI): While GPT-5.5 dominates in creative writing and a broad knowledge base, Fugu shows higher accuracy in tasks requiring strict logical sequencing.
• Claude 5 (Anthropic): Claude is known for its safety and nuance in language. However, in tests focused on complex programming (coding benchmarks), Fugu achieves results that are several percent higher due to its ability to utilize specialized coding agents.
• Gemini (Google): Google relies on integration into the entire ecosystem, but Fugu may be more attractive to developers due to its modularity and lower operating costs for individual tasks.

Practical impact: What does this mean for Czech companies and users?

You might ask: "Why should I care about a Japanese startup when I use ChatGPT?" The answer lies in efficiency and costs. If your company implements AI, you won't want to pay for a huge model for every little thing (e.g., email sorting). A Fugu-type system allows you to use "surgically precise" tools that are cheaper and faster.

Availability in the Czech Republic and Czech language support: Currently, Sakana AI primarily offers its technologies to research teams and partners via API. For a Czech user, this means that you won't see Fugu as a standalone website like chatgpt.com, but you will likely find it as part of cloud platforms or specialized developer tools. What is crucial for us in Europe is the ability of these systems to work with smaller languages. Sakana AI has previously focused on efficient training, which suggests that their systems can be very well adaptable even for Czech without the need for huge datasets in our language.

Regulation and EU AI Act: As a Japanese company, Sakana AI must comply with the strict standards of the EU AI Act when operating in the European market. The modularity of the Fugu system can be an advantage in this regard – it is easier to control and audit individual parts of the system (agents) than one opaque "black box" of a huge model. This increases the transparency that European regulation requires.

Price and implementation

Unlike a ChatGPT Plus subscription for 20 USD per month, with Fugu-type systems, we will likely work with a pay-per-token model (payment for text used). Given the efficiency of the architecture, it can be expected that the costs for complex reasoning will be lower than with OpenAI's monolithic models. For Czech startups, this opens the way to building their own applications with high intelligence, while maintaining reasonable operating costs in euros or crowns.