Sakana Namazu

Namazu is the model Sakana AI had been running inside its Sakana Chat assistant before opening it as an API on 3 August 2026. It is openly declared to be someone else's skeleton: the company builds it on Kimi K2.6, the open model published by China's Moonshot AI, and adds in-house data to fit Japanese and Japanese office practice, keigo and business documents included, plus tuning meant to cut needless refusals and to flatten political bias. Web search and code execution are built in rather than bolted on, and the endpoint speaks the OpenAI dialect, so Sakana's pitch is that switching costs one changed base_url. The vendor positions it between two bad options for Japanese firms: frontier models that cost too much, and raw open models that raise quality and data-handling doubts. Sakana published most benchmark work as a single image, so this profile carries only what the text itself states: on the company's internal FairPoliticsQA, which measures how far answers lean towards one country's values, Namazu moved from 34.10% to 56.30% against its own base model, and it also beat that base on JFBench, the Japanese instruction-following benchmark from Preferred Networks, and on an in-house Japanese-English translation task. Reasoning scores (AIME26, MMLU-Pro, LiveCodeBench v6) are described as holding the base model's level rather than exceeding it. Billing is pay-as-you-go only, with no subscription tier: $0.95 per million input tokens, $4.00 output, $0.15 cached input, thinking tokens charged as output, web search $7.00 per thousand calls and code execution $0.12 per hour of sandbox session. The model is not offered in the EU/EEA, the United Kingdom or Switzerland while the company works towards GDPR compliance.

▸Videos

No videos yet.

▸Reviews