Sakana AI Releases Fugu Ultra v1.1, Claims It Outperforms Fable 5 Without Including It in the Model Pool
Key Takeaways
- •Sakana AI claims Fugu Ultra v1.1 delivers performance gains of up to 7.9 points over the previous release, with the largest improvements on ProgramBench and TerminalBench 2.1 benchmarks.
- •All performance figures for Fugu v1.1 come from Sakana's own internal testing, and no independent verification has been made public.
- •Pricing for Fugu remains at five dollars per million input tokens and thirty dollars per million output tokens, unchanged from the prior version.
- •The v1.1 update introduces a Claude Code-compatible endpoint, allowing developers to use Fugu as a backend substitute without changing their existing terminal workflows.
- •Sakana AI continues to exclude users in the European Union and European Economic Area, citing GDPR compliance and EU-specific regulatory concerns.
Sakana AI has released Fugu Ultra v1.1, an updated version of its AI model router that distributes each user query across a pool of publicly available top-tier large language models. Model routing — dynamically selecting the best-performing model for each individual query rather than relying on a single model — has emerged as a practical strategy for developers seeking to balance cost, latency, and output quality as the number of competitive frontier models grows. The Tokyo-based company claims the update delivers performance gains of up to 7.9 points over the previous v1.0 release, with the most significant improvements observed on ProgramBench and TerminalBench 2.1 benchmarks.
According to Sakana, Fugu v1.1 now reportedly outperforms Fable 5 on certain benchmarks — even though Fable 5 is not included in the router's model selection pool. All performance figures cited come from Sakana's own internal testing, and no independent verification has been published at this time.
Pricing for the service remains unchanged at $5 per million input tokens and $30 per million output tokens. Sakana states that it takes approximately two weeks of training and evaluation before a newly released top-tier model is added to the router's available pool. The underlying architecture is detailed in the technical report.
The v1.1 update also introduces a Claude Code-compatible endpoint, allowing developers to call Fugu directly from the terminal. Claude Code is Anthropic's command-line coding assistant, and compatibility with its API format means developers already using the tool can substitute Fugu as a backend without modifying their workflows. Since its initial launch, Fugu has been accessible through platforms including OpenRouter and Vercel.
The first version of Fugu received a lukewarm reception from some users and critics, who cited concerns over high token consumption, slower response speeds, and underwhelming output quality.
Sakana AI continues to exclude users in the European Union and the European Economic Area, citing GDPR compliance concerns and what the company describes as "EU-specific regulations."
Further details are available on the official Fugu page.