On August 23, OpenRouter, without a launch event or technical report, released a model codenamed Ox Alpha with a ninja emoji, with the full identifier stealth/ox-alpha. The model is described as a frontier model for efficient coding, long-horizon agent tasks, and real production environments, featuring a context window of over 1 million tokens, a maximum output of 130,000 tokens, and support for text, image, and video inputs, but text-only outputs.
Ox Alpha supports tool calling, tool_choice, and JSON structured outputs, but not strict JSON Schema constraints. These features indicate it is not a simple chat or code completion model, but an execution model designed for coding agents, emphasizing sustained agentic work.
OpenRouter has not published any benchmark scores, such as SWE-bench or Terminal-Bench, so there is no public evidence of whether its performance reaches the top tier. Model parameters, training data, architecture, and development team are all confidential, and OpenRouter acts only as a router, with the model provided by an anonymous third-party vendor.