Sakana says its model router now beats a model it does not even use
Fugu Ultra v1.1 spreads each question across a pool of top models instead of relying on one. Sakana claims gains of up to 7.9 points and a win over Claude Fable 5. Nobody has verified that yet, and Europe still cannot use it.
Most AI products ask you to pick a model. Sakana AI’s Fugu Ultra asks a different question: why pick one at all? It is a router, a layer that sits in front of several top-tier models, decides which one should handle a given request, and stitches the results together. Sakana has now released version 1.1, and the claim attached to it is bold. The company says the router scores up to 7.9 points higher than version 1.0, with the biggest jumps on programming and terminal tasks, and that it beats Anthropic’s Claude Fable 5 even though Fable 5 is not in the pool of models it draws from.
Take that claim carefully. Every number comes from Sakana itself, and no independent testing exists yet. Benchmarks published by the company that built the thing are a starting point for a conversation, not a verdict. The architecture is described in a technical report on arXiv if you want to check the reasoning yourself.
Pricing stays where it was: $5 per million input tokens and $30 per million output tokens. Tokens are the small chunks of text a model reads and writes, and you pay per token, so output at $30 is on the expensive side. Sakana says it takes roughly two weeks of training and evaluation before a newly released top model can be added to the pool, which is a quiet admission that a router is always a step behind the frontier it routes to. The update also adds an endpoint compatible with Claude Code, so developers can call Fugu straight from the terminal, and the service remains available through platforms like OpenRouter and Vercel.
The idea behind routing is genuinely interesting. Different models are good at different things, and their weaknesses do not overlap neatly, so a system that picks the right one per task can in principle beat any single model. That is the “collective intelligence” pitch. The catch is cost and speed: the first version of Fugu got a lukewarm reception precisely because critics found it burned a lot of tokens, ran slowly, and did not deliver enough to justify either. Version 1.1 is Sakana’s answer to that, and whether it lands will depend on tests the company did not run.
One detail that matters for readers in Germany, Austria, and Switzerland: Sakana still does not serve the EU or the EEA, citing the GDPR and “EU-specific regulations.” So for a large part of this site’s audience, this is a product you can read about but not use.
What this means for you: If you are in the EU, nothing yet. If you are elsewhere and paying serious money for API calls, routers are worth watching as a category, but treat vendor benchmarks as marketing until somebody independent repeats them. The broader lesson is more portable: the assumption that you must commit to one model provider is starting to loosen. Whether the middle layer that replaces that choice is worth its own markup is still an open question.
Sources
A German open model had test answers in its training data, and openness is how we know
The consortium behind Soofi S admitted that questions from the GPQA science benchmark leaked into its training set. The community spotted it because the data was public. The team dropped the benchmark and recalculated everything.
