Tokyo-based Sakana AI has introduced Fugu and Fugu Ultra, which the company describes not as another large language model but as a multi-agent orchestration system: software that analyzes a task, splits it into parts, assigns each part to whichever AI model is best suited to it and assembles the results into a single answer.
The release is a bet that the next phase of competition in AI will be less about building the single most capable model and more about coordinating many of them.
How Fugu works
Fugu does not do most of the work itself. When a user submits a complex question or a long multi-step task, the system first analyzes the problem, breaks it into subtasks, selects an appropriate model for each one, collects and checks the outputs and then integrates them into a coherent response. Where a conventional chatbot approaches a problem with one mind, Fugu assembles a team of specialists behind a single API.
Sakana is offering two tiers. Fugu is aimed at everyday use, including chat, coding and general tasks. Fugu Ultra is designed for difficult, multi-stage problems such as scientific research, cybersecurity analysis, patent review and extended research projects.
Benchmark claims
The company’s benchmark results have drawn the most attention. According to Sakana’s technical report, Fugu Ultra outperformed Claude Fable 5, GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.8 on LiveCodeBench, a coding benchmark, and posted strong results on scientific reasoning and other complex evaluations.
The advantage was not universal. On SWE-Bench Pro, which measures long-running software engineering tasks, Claude Fable 5 remained ahead. Sakana attributes this to Fable’s design as a long-horizon agent, while Fugu’s strength lies in pooling the capabilities of several models on hard, bounded problems.
Demonstrations
Sakana also published a series of demonstrations. In one, the company says Fugu Ultra ran autonomously for about 14 hours conducting machine learning research, carrying out more than 100 experiments and identifying improvements to the training of a small GPT-style model. In another, it was asked to make investment decisions from historical financial data in a simulated environment and, according to the company, outperformed other leading models. Sakana also reports strong results on blindfold chess, Rubik’s Cube solving and mechanical engineering design tasks.
Sovereignty as a feature
One idea the company emphasizes is what it calls AI sovereignty. Because Fugu sits above individual models rather than depending on one, it can keep working if a particular model becomes unavailable, is restricted in a given country or has its access curtailed, by routing tasks to other models instead. For governments and enterprises wary of relying on a single vendor, that resilience may be as much of a selling point as raw performance.
Reasons for caution
All of these results come from Sakana’s own evaluations. Several of the benchmarks are widely used, but the company’s claims have not yet been independently reproduced, and orchestration systems can be sensitive to which underlying models are available and how they are configured. A clearer picture of Fugu’s real-world standing will emerge once third-party researchers test it.
Even so, the approach reflects a shift already visible across the industry. After several years in which attention centered on building ever larger models, the question of how to combine existing models into effective systems is becoming a competitive frontier of its own.
