Japanese startup Sakana AI has released version 1.1 of its Fugu Ultra router, providing a timely look at how architectural trends are shifting in the AI industry. While tech giants compete over the parameter counts of monolithic models, the Tokyo-based team is selling the role of the conductor rather than "intelligence" itself. According to the company's report, the updated router gained 7.9 performance points over the first version, showing its most impressive results on the ProgramBench and TerminalBench 2.1 benchmarks. Ironically, Sakana AI claims its system outperforms Fable 5, even though the latter is not currently included in the router's pool of available models.
The business logic here is simple and pragmatic: why bet everything on a single "universal" LLM when you can manage an ensemble? At $5 per million input tokens and $30 per million output tokens, Sakana offers efficient load balancing. The router decides whether to send a coding request to a specialized model or delegate complex reasoning to a recognized market leader.
As Sakana AI explains, version 1.1 now supports a Claude Code-compatible endpoint, allowing users to run Fugu directly from the terminal or via familiar platforms like OpenRouter and Vercel.
For CTOs and architects, this is a clear signal: direct subscriptions to proprietary giants are becoming a questionable luxury if a "smart layer" can optimize cost and quality better than in-house engineers. However, this technological honey has its share of tar.
The primary risk for dynamic production systems is the two-week lag Sakana AI requires to fine-tune and evaluate new top-tier models before adding them to the pool. In an industry where leaders change monthly, such sluggishness could prove fatal. The first version of Fugu faced criticism for excessive token consumption and slow processing speeds. Independent benchmarks for version 1.1 are still non-existent, and the company has pointedly refused to serve EU clients due to GDPR regulatory hurdles.
For now, the venture looks like an attempt to build a business by reselling other people's compute power. The ultimate question is whether a proprietary router can maintain its edge when its success depends entirely on the very third-party models it seeks to abstract away from the user.