Japanese lab Sakana AI launched two new products on September 11, 2026: Fugu Max and Fugu Ultra v2, which aren't traditional language models but orchestrators — systems trained to split each task across an interchangeable pool of models and assemble the final answer from their outputs.
Instead of answering each query directly, Fugu receives the request through a single API and decides, task by task, which model or models in its internal "pool" should handle it, then combines their outputs into a single response. It's a different bet from most labs, who compete by shipping their own biggest, most capable model.
Fugu Max costs $2 per million input tokens and $6 per million output tokens. Sakana claims this represents 40% to 60% less than the output pricing of Claude Sonnet 5 and Kimi K3. The model pool Fugu Max orchestrates includes open-weight and specialized models, including NVIDIA's Nemotron family.
The most striking detail of the launch is that Fugu Ultra v2, Sakana's highest-capability product, no longer includes Claude Fable 5, Claude Fable 5.1, or GPT-6 Astra in its pool — three of the most powerful systems available today — yet Sakana claims it still beats all three on several of its own benchmarks. On "Chartography," Fugu Ultra v2 scores 48.3 points versus 27.3 for Claude Opus 5; on DeepSWE it reaches 74.3 points.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel