Japanese company Sakana whose Fugu AI model matched Claude Fable 5's performance shares a message for all countries banned from using Anthropic's most powerful AI modelsJapan's Sakana AI has built a workaround for one of the AI industry's newest problems: what happens when the US government cuts off access to the world's best models. Its answer is Fugu, a multi-model orchestration system launched on June 22 that the company says matches Anthropic's restricted Claude Fable 5 and Mythos models on benchmarks -- without using either of them. The system pools together publicly available models like Google's Gemini 3.1 Pro, OpenAI's GPT-5.5, and Claude Opus 4.8, then dynamically picks and combines them for each task.The timing is no accident. Fugu arrives days after a US export control order forced Anthropic to pull its top models from global use on June 12, locking out entire countries overnight. Sakana's message to those left stranded is blunt: you don't need the banned models to get banned-model performance. CEO David Ha is pitching orchestration not just as a clever engineering trick but as a hedge against the kind of single-vendor dependence that can collapse the moment Washington changes its mind.Fugu, Japanese for pufferfish, is itself a language model trained to call other LLMs from an agent pool, including instances of itself recursively. You send a request to one endpoint, and Fugu decides everything internally -- model selection, delegation, verification, and synthesis. It solves simple tasks on its own and assembles a team of expert models when a problem needs more. None of that complexity ever touches your code.Sakana offers two versions through a single OpenAI-compatible API. The base Fugu balances performance with low latency for everyday work and slots into tools like Codex, and you can opt specific agents out of its pool for compliance reasons. Fugu Ultra is the flagship, tuned for hard, multi-step jobs like AI research, cybersecurity analysis, and patent investigations. On Sakana's benchmarks, Fugu Ultra posted 73.7 on SWE-Bench Pro, ahead of Opus 4.8 at 69.2, and 95.5 on the GPQA-D science test. The catch, critics note, is that an orchestrator can only be as good as the models it can reach -- and Fugu still trails the restricted Fable 5 on several tougher tasks.For Ha, the export controls are the whole point. He believes the strongest AI systems won't be isolated monoliths but collective intelligences, much like human progress itself. "Human intelligence is fundamentally a collective intelligence," he wrote on X, arguing that the future belongs to systems that learn how to coordinate. Fugu, he says, is a major step toward that. Relying on a single company's model for national infrastructure is a massive risk, he argues, and recent export controls prove access to top models can vanish overnight. Collective intelligence is his answer -- a practical hedge against that concentration of power, with Fugu routing around vendor restrictions through an entirely swappable agent pool. "By orchestrating the world's models, we are delivering the resilient blueprint required for AI sovereignty," he wrote, crediting his Tokyo team for shipping it.Not everyone is convinced. Prime Intellect engineer Elie Bakouch called it a closed orchestrator built on closed models, arguing users now control even less than before. Pricing runs steep too, with Fugu Ultra matching GPT-5.5 at $35 per million tokens combined. Fugu is live everywhere except the EU and EEA, where Sakana is still squaring its routing with GDPR.