Sakana AI

DataNorth reported on September 11, 2026 that Sakana AI released Fugu Max and Fugu Ultra v2 as two new variants of the Fugu orchestrator architecture. Fugu Max is priced at $2 per million input tokens and $6 per million output tokens, with a $0.25 per million cached-input rate, while Fugu Ultra v2 is priced at $5 per m Fugu Ultra v2 targets the best possible answer on hard, multi-step work rather than cost optimization, keeping the same orchestrator architecture but tuning for deeper agent pools and higher quality, at the price of greater latency and cost. The DataNorth piece notes that Sakana frames the swappable model pool as prote

An independent analysis published August 11, 2026, synthesizes two Sakana AI announcements into a "sovereign-by-tuning" strategy. Namazu, released August 3, 2026, is described by Sakana as a fine-tune of Moonshot AI's open-weight Kimi K2.6, not a Sakana-pretrained model, and is priced at $0.95 per 1M input tokens, $4.0 One week later on August 10, Sakana announced it had retrained Fugu's small "conductor" model on a Google Gemma 4 base as a portability test, not a production swap, demonstrating the orchestration layer can run over different open-weight backbones. A material caveat for developers: Namazu is not available in the EU/EEA

An independent technical review examines Sakana AI's "Fugu," a multi-agent orchestration system released in June 2026 and shipped as a single OpenAI-compatible API. Rather than acting as a single model, Fugu assigns roles (a Thinker, Workers, and a Verifier) to a pool of LLMs and routes queries among them, with the aut The review flags significant caveats on the headline scorecard: the numbers are self-measured by Sakana against vendor-reported baselines, Fugu loses at least one row in Sakana's own table, the public demo hides competing models behind shuffled labels, and the large token and latency overhead of orchestrating multiple