Fugu Ultra v1.1 benchmarks: Sakana says router tops Fable 5
Sakana AI says its updated model router improves on v1.0 and beats Fable 5, but the benchmark results have not been independently verified.
By Colin Brandt · Enterprise Reporter
· 3 min read
Sakana AI has released Fugu Ultra v1.1, a revised version of its AI model router, and says the Fugu Ultra v1.1 benchmarks now put it ahead of Anthropic’s Fable 5 on most tests. The company kept pricing unchanged at $5 per million input tokens and $30 per million output tokens, while making a performance claim that matters for buyers comparing routed model access with direct frontier-model APIs.
The headline claim is specific but still company-reported: Sakana says v1.1 improves on Fugu v1.0 by as much as 7.9 points, with the largest gains on ProgramBench and TerminalBench 2.1. Sakana also says Fugu v1.1 outperforms Fable 5 even though Fable 5 is not one of the models available for the router to choose from. No independent verification of those benchmark results has been published.
What is Fugu Ultra v1.1?
Fugu is Sakana AI’s router for sending each user request to a pool of publicly available top-tier models, rather than relying on a single underlying model. In practical terms, the product is selling model selection as the layer of differentiation: the system decides which model should handle a given query, with the goal of improving output quality across varied tasks.
Sakana describes the architecture in a technical report and says adding a newly released top-tier model to Fugu’s pool requires about two weeks of training and evaluation. That delay is an operational detail worth watching because routing products depend on staying current with the strongest available models while also proving that the routing logic adds value beyond model aggregation.
The update also adds a Claude Code-compatible endpoint, which Sakana says allows developers to call Fugu from the terminal. Since its launch, Fugu has also been available through platforms including OpenRouter and Vercel.
What changed from the first Fugu release?
The first version of Fugu drew a restrained response. Critics pointed to high token consumption, slow performance and weak results. Sakana’s v1.1 announcement is therefore less a routine model update than an attempt to reset the product’s technical standing against the category’s best-known proprietary systems.
The Fable 5 comparison is the most notable part of the release because it frames Fugu as a way to match or exceed a leading model without directly routing to it. That is also the part that needs outside testing. Router benchmarks can be sensitive to task mix, model availability, prompting and cost assumptions, so the company’s numbers are useful as a claim to test rather than as settled evidence.
Sakana has not changed Fugu’s regional availability. The company still does not serve customers in the EU or EEA, citing GDPR and other EU-specific regulations. That limits the immediate addressable market for European developers and enterprises evaluating whether a router can reduce dependence on a single model provider.
For AI infrastructure buyers, Fugu v1.1 adds another data point in the shift from choosing one foundation model to buying orchestration across several. Sakana’s case now rests on whether independent users see the same benchmark improvements, and whether the latency, token usage and reliability concerns from the first release have been materially reduced.
This story draws on original reporting from The Decoder.
