Tokyo-based Sakana AI has released Fugu, a multi-agent orchestration system that lets users tap into a pool of language models through a single API. The flagship tier, Fugu Ultra, claims to match the performance of Anthropic’s Fable and Mythos models — the same models that were yanked offline in June under a U.S. export control directive.
Why Anthropic’s models disappeared
On June 12, Anthropic suspended access to Fable 5 and Mythos 5 after a U.S. national security directive barred foreign nationals from using the models. The move left developers and researchers who relied on those models scrambling for alternatives. Sakana AI’s Fugu system is designed as a hedge against exactly that kind of vendor lock-in. By orchestrating a pool of swappable models — including open-weight and commercial ones — Fugu can route tasks to whatever model is available and best suited for the job.
How Fugu works
Fugu is not a single large language model. It’s a system that calls other models in an agent pool, can invoke itself recursively, and handles selection, delegation, and synthesis internally. Developers interact with it through a single OpenAI-compatible API. The system ships in two tiers: Fugu for everyday tasks with lower latency, and Fugu Ultra for complex work like AI research and cybersecurity analysis.
Benchmark results
Sakana ran Fugu Ultra against Fable 5, Mythos Preview, Gemini 3.1 Pro, GPT 5.5, and Opus 4.8 across eight tests. Fugu Ultra led four of them: LiveCodeBench at 93.2 and GPQA-D at 95.5. Fable 5 still topped three benchmarks — SWEBench Pro (80.0 vs Fugu Ultra’s 73.7), SciCode, and Humanity’s Last Exam. Opus 4.8 took the lead on CTI-REALM.
In an autonomous machine learning test, Fugu Ultra ran more than 100 experiments in 14 hours on a single H100 GPU, achieving the best mean performance (0.9774) and the best single run (0.9748).
Broader market context
The launch comes as AI model access grows more volatile. While Sakana pushes an orchestration approach, other players are moving in different directions. India’s Sarvam AI, for instance, recently hit a $1.5 billion valuation by building its own models tailored for Indian languages. For now, Sakana is betting that the ability to swap models in and out of a pipeline — without rewriting code — will become a must-have for teams that can’t afford to lose access overnight.
The question that remains is whether Fugu’s performance will hold up as Anthropic and other labs release updated versions of their restricted models, and whether the orchestration layer itself becomes a new point of failure.
