While media attention recently has been focused on Anthropic and OpenAI, a new AI announcement from outside Silicon Valley quietly upstaged them on June 22nd. Fugu is a new AI system from a Japanese AI lab called Sakana, built around a large language model and an orchestration layer. On the challenging SWE-Bench Pro benchmark, Sakana reported that Fugu beat Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, as well as Gemini 3.1 Pro. Fugu Ultra also beat Anthropic’s Fable 5 on LiveCodeBench (scores were Fugu Ultra: 93.2, Fugu: 92.9, Fable: 89.8). Further, it beat Mythos Preview on GPQA-Diamond, a test of 198 graduate-level multiple-choice questions in biology, physics, and chemistry (scores were Fugu Ultra: 95.5, Fugu: 95.5, Mythos Preview: 94.6).
Sakana (“fish” in Japanese) was established in 2023 in Tokyo. Its co-founders have some serious credentials. Llion Jones was co-author of the widely cited “Attention Is All You Need” research paper while at Google. This paper unveiled the transformer architecture that sits at the heart of all current large language models. David Ha, the company’s CEO, was formerly a Google Brain researcher. The company has funding from global venture capital firms, NVIDIA, and Japanese banks.
Instead of competing to build an ever-larger monolithic LLM, Sakana works on systems where models and agents work together collaboratively. An analogy might be that Fugu is the conductor of an orchestra of other LLMs. It is a co-ordination system. In response to a prompt, Fugu decides what kind of problem it is, and which resources are best to address it. A simple question may be answered directly, while a coding problem may be routed to a model that is particularly good at coding. It can call on other LLMs and even recursively call other copies of itself. It assigns various roles like “planner”, “solver”, “verifier” and “critic”, calls these resources, combines their answers, evaluates the outputs, resolves inconsistencies, and produces a consolidated answer. In fact, there are two new models: Fugu Ultra does multi-step reasoning, while Fugu is cheaper and faster and is used for general tasks such as chat.
Some caveats need to be pointed out. These are vendor reported benchmark results, and benchmarks themselves are tricky things. There are many competing benchmarks, and vendors tend to highlight ones that their practical model does well at. Nonetheless, the Fugu results seem impressive, and they show that there is not just one way to go about LLM architecture. Fugu is an LLM, but Sakana has not published the size of the models in terms of parameters, though to be fair neither does Anthropic or OpenAI. It is likely that the Fugu model itself is fairly small, as it relies on other models to do the heavy lifting, while it focuses on orchestrating these models. The idea is not unique to Sakana. There is also the “LLM Council” concept of AI scientist Andrej Karpathy, and OpenRouter’s Fusion, an API gateway to other models. However, Sakana appears to be the first vendor to have designed an LLM specifically trained on task delegation to act as an orchestrator of other models. It should also be noted that Fugu Ultra has considerable latency, where answers may take many minutes to be produced.
What is particularly interesting is that Sakana takes quite a different approach to the leading US frontier labs, which are in a race to build ever larger models based on ever more training data. The Sakana approach appears to show that careful orchestration of a swarm of models can outperform a monolithic LLM, at least assuming that the benchmark results stand up to scrutiny. Having a layer above core LLMs may have other advantages. Just recently the US government blocked the release of Anthropic’s Claude Fable, showing that geopolitics can be an important factor in AI these days. An orchestration approach that does not depend on any one LLM frontier model may be quite appealing to countries and organisations that are nervous about lock-in to any particular model, especially if that model may be blocked on a whim of the US government.
Along with the increasing capability of open weight models like GTM 5.2, DeepSeek V4, Qwen 3.6 and others, the release of Fugu opens up the world of high-performance AI models beyond the US frontier labs. The future of AI may not belong exclusively to ever-larger foundation models. It may also belong to systems that learn how to coordinate them effectively.
