Sakana AI has launched Fugu Ultra v1.1, an updated version of its AI model router, claiming significant performance improvements over its predecessor. This release comes with the assertion that Fugu Ultra v1.1 can outperform Anthropic's Fable 5 on several benchmarks, despite Fable 5 not being part of the selection pool used by Fugu.
Fugu Ultra v1.1 is designed to distribute queries across a pool of publicly available top-tier models, offering performance gains of up to 7.9 points over the previous version, particularly on ProgramBench and TerminalBench 2.1. The pricing remains consistent at $5 per million input tokens and $30 per million output tokens. Additionally, the update introduces a Claude Code-compatible endpoint, allowing users to call Fugu directly from the terminal. Fugu has been accessible on platforms such as OpenRouter and Vercel since its initial release.
The technical approach of Fugu Ultra v1.1 involves a two-week training and evaluation period before new models are added to its pool. While Sakana has released a technical report detailing the architecture, independent verification of its performance claims is currently unavailable. The previous version of Fugu faced criticism for high token usage, slow speed, and subpar results, but Sakana aims to address these issues with the latest update.
Fugu Ultra v1.1 is targeted at developers and enterprises looking for a robust AI model routing solution. Its use cases include optimizing query distribution across multiple models to enhance performance in a variety of applications.
Work implications: Fugu Ultra v1.1 could enhance workflows for developers and enterprises by optimizing model selection and improving response times, potentially benefiting roles involved in AI deployment and model management.
Originally reported by the-decoder.com