A new artificial intelligence model, dubbed “Ox Alpha,” has recently appeared on the OpenRouter platform, generating significant buzz due to its advanced capabilities and the complete anonymity of its creators. Described as a “stealth model,” Ox Alpha is reportedly developed and operated by an undisclosed third-party provider, who has chosen to remain anonymous during its preview phase.
The model's emergence has caught the attention of prominent figures, including Stripe CEO Patrick Collison, who described Ox Alpha as “very impressive.” Speculation regarding its origins largely points towards China, with AI analyst Andrew Curran suggesting it might be a new iteration from the GLM AI model family.
What is Ox Alpha?
Ox Alpha is characterized as a powerful, long-horizon multimodal AI model. This means it can efficiently process and interpret various types of data simultaneously, including text, images, and potentially video, to tackle complex tasks. A key feature is its reported 1-million-token context window, allowing it to handle an exceptionally large volume of information within a single interaction. This capability positions it for demanding applications suchs as intricate coding, sustained agentic tasks, and production-level workloads.
Performance and Potential Creators
Early, vendor-presented comparisons highlight Ox Alpha's strong performance. In an independent 10-task coding test, the model successfully solved 8 tasks. For context, competitors like Fable 5 achieved 65%, GLM-5 62%, GPT-5 52%, and Grok 4 62%. It's important to note that these are not broadly established benchmarks but offer an initial glimpse into its potential.
The identity of Ox Alpha's developer remains a central mystery. While the community is abuzz with theories, no public confirmation has been made. Leading theories suggest it could be an unreleased GLM model, possibly a GLM-5.3 variant from Z.ai, or even Microsoft's unreleased MAI model. As its operator, long-term pricing, service guarantees, benchmark performance, and data-handling arrangements are still unclear, the model is currently advised to be treated as an experimental preview.