Sakana's Fugu Ultra v1.1 claims to beat Claude's Fable 5 without using it

Sakana AI has released Fugu Ultra v1.1, an updated version of its AI router—a system that distributes incoming queries across a curated pool of publicly available top-tier models rather than relying on a single monolithic model. The company claims the update delivers performance gains of up to 7.9 points relative to v1.0, with the most significant improvements appearing on ProgramBench and TerminalBench 2.1. In a claim that could attract attention, Sakana reports that Fugu Ultra v1.1 outperforms Anthropic's Fable 5 on most benchmarks, even though Fable 5 is not included in the router's model pool.

It is important to note that these performance numbers come entirely from Sakana and have not yet been verified by independent third parties. The pricing model for Fugu Ultra v1.1 remains unchanged from prior versions: $5 per million input tokens and $30 per million output tokens. Sakana indicates that when a new top-tier model becomes available, it undergoes about two weeks of training and evaluation before being integrated into Fugu's selection pool. The technical details of this update are documented in Sakana's technical report. In addition to performance improvements, v1.1 introduces a Claude Code-compatible endpoint, allowing developers to call Fugu directly from the terminal.

Fugu has been accessible via platforms such as OpenRouter and Vercel since its launch. The earlier version of Fugu received mixed to negative feedback from users and critics, who highlighted concerns about high token consumption, slow response speeds, and weak results. A regional limitation applies: Sakana does not currently serve the EU or EEA, a restriction the company attributes to GDPR and other EU-specific regulatory requirements.

Fugu Ultra v1.1 represents Sakana's attempt to position AI routing—intelligent distribution of queries across multiple models—as a competitive alternative to building or fine-tuning a single large model. The claimed 7.9-point performance gain over v1.0, particularly on ProgramBench and TerminalBench 2.1, suggests optimization is occurring; however, the lack of independent verification is a significant caveat. The fact that Sakana claims v1.1 outperforms Fable 5 without including it in the pool could indicate either genuine architectural efficiency or that the benchmarks themselves may not be representative of real-world use.

The first Fugu version encountered skepticism due to high token usage, slow response times, and disappointing results—issues that a 7.9-point gain may address, but only if verified externally. The addition of a Claude Code-compatible endpoint signals Sakana's intent to integrate more deeply into developer workflows. The two-week evaluation window for adding new models to the pool suggests Sakana has a process to keep Fugu current as new models emerge, though this also means the pool is not static and claims about v1.1's performance may shift as models are added or removed.

No comments yet. Be the first to share your thoughts!