Overview
- An anonymous “stealth” model called Ox Alpha went live on OpenRouter on August 20 with a free limited preview that routes queries through multiple portals.
- The provider advertises a roughly 1,048,576‑token context window, multimodal input for text, images and video, and very high serving capacity that drove intense developer use.
- Early small‑sample benchmarks produced viral claims that Ox Alpha outperformed leading coding models but fuller 113‑task DeepSWE runs later showed roughly comparable scores near 63 percent.
- Independent black‑box fingerprinting of tokenizer behavior and response patterns has repeatedly matched signals from Zhipu AI’s GLM‑5.3 family but attribution remains unconfirmed.
- Published terms conflict over prompt and output retention versus training use, creating immediate privacy risk for teams that feed private code into the model and raising export‑control and procurement concerns.