Sakana AI has launched Fugu, a multi-agent orchestration system, with its flagship Fugu Ultra claiming to match the performance of Anthropic’s export-restricted Fable and Mythos models.

The Tokyo-based startup positions Fugu as a hedge against vendor lock-in, arguing that orchestrating a pool of swappable models lets users sidestep the restrictions that recently pulled Fable and Mythos offline.

How Sakana Fugu Works and Why It Matters

Fugu is a large language model (LLM) that calls other models in an agent pool. It can even invoke itself recursively. Users send one request, and the system handles selection, delegation, and synthesis internally.

The product ships in two tiers via a single OpenAI-compatible API. 

• Fugu targets everyday work with lower latency. 
• The flagship model Fugu Ultra coordinates a larger agent pool for more complex tasks, such as AI research and cybersecurity analysis.

The launch follows Anthropic’s June 12 suspension of access to Fable 5 and Mythos 5. A US export control directive cited national security and barred access for foreign nationals. 

Sakana’s pitch taps the sovereign AI debate, the idea that nations want control over the models they depend on. 

The team argued that any government or business that relies on a single provider’s AI for essential systems “is a material vulnerability.” Sakana said that the threat has shifted from theory to something already playing out.

“As we have seen with recent export controls imposed on models like Fable and Mythos, access can disappear overnight. Collective intelligence is the practical hedge against this concentration of power. Because Fugu orchestrates an underlying pool of swappable agents, it simply routes around vendor restrictions. By orchestrating the world’s models, we are delivering the resilient blueprint required for true AI sovereignty,” the team said.

Meanwhile, progress is also being made in other nations. India’s Sarvam AI recently reached a valuation of $1.5 billion.

How Fugu Ultra Stacks Up Against Fable, Mythos, and Frontier Rivals

Sakana benchmarked Fugu against Fable 5, Mythos Preview, Gemini 3.1 Pro, GPT 5.5, and Opus 4.8. The results are mixed rather than a clean sweep. Fugu Ultra led four of eight tests, including LiveCodeBench at 93.2 and GPQA-D at 95.5.

Fable 5 came out on top in three benchmarks. It topped SWEBench Pro at 80.0, well above Fugu Ultra’s 73.7. Fable 5 also led on SciCode and Humanity’s Last Exam, while Opus 4.8 edged ahead on CTI-REALM.

Against Mythos Preview, Fugu Ultra won each shared benchmark, but only by sub-point margins. Beyond the charts, Sakana also ran task demonstrations.

In an autonomous machine learning test, Fugu Ultra ran more than 100 experiments over 14 hours on a single H100 GPU.

“Fugu Ultra (bold red) finished with the best mean performance (0.9774). Fugu Ultra also achieved the best single run of the entire experiment (0.9748), leading every single baseline,” the team said.

The other demos covered trading, chess, and design. Overall, Fugu joins a growing effort by nations and firms to build alternatives as access to frontier models tightens.

Source: BeInCrypto