Top items
- Sakana AI releases Fugu Max and Fugu Ultra v2, learned orchestrators that route queries across a pool of open-weight and specialist models and reportedly beat several frontier models on benchmarks.
Company & product developments
Sakana AI launches Fugu Max and Fugu Ultra v2 model orchestrators (sakana.ai, via AI Weekly Alerts). On September 11, 2026, Sakana AI released two new products — Fugu Max and Fugu Ultra v2 — built as a single learned orchestrator that routes incoming queries across a pool of open-weight and specialist models, all exposed behind one OpenAI-compatible API. The design pitch is that rather than relying on a single monolithic frontier model, the orchestrator learns to dispatch each query to whichever underlying model is best suited, and presents the result through a standard interface. On pricing, Fugu Max is set at $2 per million input tokens and $6 per million output tokens, which Sakana positions at 40–60% below competing frontier offerings named as Sonnet 5 and GPT 5.6 Terra. On performance, Fugu Max is claimed to rank best on six of ten benchmarks cited, including Terminal Bench 2.1 and GPQA Diamond. The higher-end Fugu Ultra v2 reportedly scores 48.3 on Chartography and 74.3 on DeepSWE — and notably, Sakana says it achieves this without including Fable 5, Fable 5.1, or GPT-6 Astra in its agent/model pool, implying the results come from orchestrating a more modest set of underlying models. The combination of lower cost and competitive benchmark performance is the core of Sakana’s claim that these orchestrators can beat frontier models.