Fugu Ultra v2 is Sakana AI's high-performance model, but it works differently than typical AI systems. Instead of being a single large model, it's a routing system—a language model trained to distribute tasks across a fixed pool of specialized open-source models and call itself recursively when needed. The architecture prioritizes complex reasoning, multi-step research tasks, and full-stack software development. It doesn't depend on proprietary frontier models, which is a meaningful constraint that affects both capability and cost structure. You can configure reasoning effort at three levels (high, xhigh, max), use function calling and structured outputs, feed it images and PDFs, and run web searches directly. Pricing runs $5 per million input tokens and $30 per million output tokens—expensive on the output side, which matters for verbose responses. Cache reads cost $0.50 per million tokens, and web searches run $10 per 1,000 calls. The context window maxes out at 1 million tokens, with up to 128,000 completion tokens per response. Requests over 272K tokens get billed at a higher rate. In practice, customers pay less than the listed rates thanks to caching and volume discounts. Performance metrics show 31 tokens per second throughput, 10.6 second latency at the median, and 87.47% availability over the past three days—solid but not exceptional. The model launched September 11, 2026, with an August 28 knowledge cutoff. Hermes Agent, an open-source self-improving AI system from Nous Research, has sent the most traffic to it so far, consuming 112 million tokens. That's a useful signal: Fugu Ultra v2 appears well-suited for autonomous agent work that benefits from task routing and recursive reasoning rather than raw model size. It's hosted through OpenRouter, which means you get their load-balancing infrastructure and can swap between providers if one fails.