Sakana AI just dropped two new models, Fugu Max and Fugu Ultra v2, but they aren’t your typical foundation models. Think of them as smart traffic controllers rather than engines. They don’t generate text from scratch; instead, they read your query and dynamically route it to the most efficient model in their pool. This architecture, built on evolutionary algorithms and reinforcement learning, aims to solve the industry’s biggest headache: paying premium prices for simple tasks.
Fugu Max is the budget-conscious workhorse. Priced at $2 per million input tokens and $6 for output, it claims to be 40% to 60% cheaper than competitors like Sonnet 5 or Kimi K3. It achieves this by widening its roster of available models, including NVIDIA’s Nemotron family, and picking the leanest option that can still get the job done. Sakana reports it tops six major benchmarks, including Terminal Bench 2.1 and SWEFish, while expanding the cost-performance Pareto frontier on seven out of ten tests. For startups burning through API credits, that efficiency gap is hard to ignore.
On the other end of the spectrum, Fugu Ultra v2 chases raw capability. It skips the usual suspects like Fable 5 and GPT-6-Astra, relying instead on a specialized pool to handle complex reasoning. The results are striking: it scored 48.3 on Chartography, nearly doubling Opus 5’s 27.3, and hit 74.3 on DeepSWE. This suggests it can handle sustained reasoning over visual and structured data better than models costing three to five times more per token. It’s a bold move to exclude top-tier proprietary models, betting that orchestration beats brute force.
Both models are live today via an OpenAI-compatible API, meaning you can swap them in with a single line of code. However, there are caveats. You can’t self-host these since no open weights are released, and the service isn’t available in the EU or EEA. While Sakana argues this reduces vendor lock-in, you’re simply swapping one black box for another. The real test won’t be benchmark scores, but whether enterprises trust a router they can’t inspect to handle their most sensitive data.
