Japan's Sakana AI claims its Fugu Ultra model matches Mythos on certain benchmarks — and reportedly beats Claude 5

Japan's Sakana AI has unveiled Fugu Ultra, a new AI model that the company says matches Mythos on certain benchmarks, with subsequent reporting indicating the system also beats Anthropic's Claude 5 on specific evaluations — a striking claim that has drawn significant attention in the global AI community.

Fugu is not a conventional general-purpose AI assistant. Sakana has positioned it as an AI agent-orchestrator: a system designed to help businesses and governments navigate the increasingly complex web of technology export controls that now govern the international movement of AI tools, chips, and related technologies. The specialised focus makes it a product of Japan's particular geopolitical and regulatory reality, where export compliance has become a critical operational burden.

The benchmark results, if they hold up to independent scrutiny, would represent a significant milestone for non-US AI development. Until now, the global frontier model race has been dominated by American labs — OpenAI, Anthropic, Google DeepMind — with Chinese players like DeepSeek representing the main external challenge. A Japanese model performing competitively on key benchmarks would diversify the competitive landscape meaningfully.

Sakana AI was founded by former Google researchers and has positioned itself as a laboratory exploring alternative approaches to AI development, often drawing on principles from natural systems and evolution. Fugu Ultra represents its most ambitious product announcement to date.

For the broader AI industry, the story raises a pointed question: as specialised, regulation-aware AI models emerge from non-US jurisdictions, what does that mean for the West's current dominance of both AI capability and AI governance norms?