In a competitive landscape where collective action often outstrips singular power, independent artificial intelligence companies are striving to bridge the gap between open-weight and proprietary frontier models. Their innovative approach emphasizes the synergy generated through collaboration rather than the dominance of a singular, powerful entity.

One notable player in this arena is Sakana AI. This Japanese company claims to have developed a sophisticated system that matches the capabilities of leading AI models, including Anthropic’s Mythos preview and OpenAI’s GPT-5.5. Their flagship product, a model orchestration API named Fugu, operates on a unique principle: it leverages a network of agents and models to complete various tasks rather than relying on a single, high-powered model.

Fugu does not fit the typical mold of an open-source large language model which is designed to undertake various functions unilaterally. Instead, Sakana AI characterizes it as an “orchestration model.” This architecture enables Fugu to operate through a multi-agent system that presents itself as a unified large language model (LLM). Upon receiving a request from users, the Fugu system intelligently determines the best approach to handle that request by coordinating the efforts of several agents. If the task at hand is particularly complex, Fugu can tap into additional models to ensure a thorough and effective response.

Sakana AI advocates that this model of collective intelligence serves as a safeguard against overreliance on any single firm. Such vulnerability can lead to significant challenges, particularly in light of recent disruptions such as the U.S. government’s decision to enforce export controls on Fable 5 and Mythos 5, which have wide-reaching implications for businesses and organizations reliant on specific AI technologies.

“The recent upheavals in the AI landscape have starkly illustrated the dangers associated with dependency on singular vendors,” noted Sakana in a recent blog post. “Organizations or even nations that rely heavily on a single company’s APIs for essential functions such as infrastructure, finance, or governance expose themselves to material risks. This is no longer a theoretical concern; it is a pressing reality.”

Sakana has introduced two versions of Fugu. The base Fugu is designed to balance performance with low latency, making it ideal for coding tasks. In contrast, Fugu Ultra is tailored for tackling complex, multi-step problems that require a high degree of accuracy and depth. Sakana claims that Fugu Ultra stands on par with other leading models like Fable 5 and Mythos Preview, particularly in rigorous benchmarks related to engineering, scientific inquiry, and reasoning.

In a market dominated by financially robust entities like Anthropic, OpenAI, and Google—each possessing vast research resources—the competition is fierce. The last few years have seen open-source initiatives making significant strides, challenging the performance capabilities of frontier models. This shift is exemplified by the emergence of DeepSeek R1, Alibaba’s Qwen models, and Mistral’s offerings. However, many open-weight models often find themselves limited due to a lack of substantial data, computational power, and extensive post-training alignment—resources that many proprietary model providers can access with ease.

Conversely, many enterprises express concerns regarding the level of model provenance, data handling, and auditability provided by open-source projects. In response, Sakana contends that the Fugu system effectively addresses these concerns. Rather than compelling enterprises to choose between existing models that fulfill their needs within open and closed ecosystems, Fugu provides a coordinated framework to maximize the capabilities of various models. This collective intelligence approach allows for more robust outcomes than those generated by a singular model alone.

According to Sakana’s internal benchmark testing, Fugu Ultra achieved a score of 93.2 on LiveCodeBench, outperforming both Fable 5, which scored 89.8, and GPT-5.5, which recorded an 85.3. For SWEBench Pro, Fugu Ultra fell short of only Fable 5. However, it is crucial to note that these scores have not been independently verified.

Initial user feedback on social media has revealed a spectrum of opinions. Wharton professor Ethan Mollick remarked on the speed of Fugu Ultra, indicating it started off slow but ultimately delivered satisfactory performance. In contrast, other individuals praised the novelty and ingenuity of Sakana’s approach in the AI field.

Aaron Levie, the CEO of Box, emphasized the burgeoning trend of applied AI products that develop agent harnesses. He posited that the concept of making this orchestration an LLM accessible to developers was highly promising. As innovation flourishes within both frontier closed and open-source models, Levie foresees a wealth of value generated for the layer capable of effectively routing the best solutions.