Sakana's Fugu router tops frontier models on 5 of 8 benchmarks
Not a single model but a learned router over open + specialist ones — OpenAI-compatible and on OpenRouter now; Max runs $2/$6, ~40% under Sonnet 5 and GPT-6.
Copy markdown
One API, a whole pool of open models
Fugu isn't a model — it's a learned orchestrator that dispatches each task to the lightest capable open or specialist model, can recursively call itself, then assembles the outputs. Fugu Max optimizes for cheapest-that-works; Ultra v2 for peak quality.
It beats frontier models it doesn't even contain
Ultra v2 posts 48.3 on Chartography (visual reasoning) vs Opus 5's 27.3 and Fable 5's 29.5, and 74.3 on DeepSWE — best or tied-best on five of eight benchmarks, with no frontier model in the pool.
Drop-in: point your OpenAI client at it
The API is OpenAI-compatible and live on OpenRouter — switch with a single parameter, no migration. Fugu Max is $2/$6 per M tokens (~40-60% under Sonnet 5 and GPT-6 Terra); Ultra v2 is $5/$30 with a 1M-token context and 128K max output.
The catch: reliability is still shaky
On OpenRouter, Ultra v2's successful-inference rate was 78.7% over 24h (86.8% over 3 days) even at 100% provider uptime — build in retries and fallbacks. Prompts over 272K tokens bill at a higher tier, and cache reads run $0.50/M.
Build this weekend: a cheap agent that punches up
Repoint an existing OpenAI-SDK agent at Fugu Max and get frontier-ish coding and reasoning at open-model prices; reach for Ultra v2's 1M context, web search, and PDF/image input to build a research agent that reads long docs without a frontier bill.