Ox Alpha appeared on OpenRouter on August 20 as a “stealth” reasoning model aimed at coding, sustained agentic work and production workloads. Its developer and operator have not been identified publicly, while OpenRouter says explicitly that it only routes requests to the model and does not own, develop or operate it.
The model offers a context window of 1,048,576 tokens, placing it among a small group of systems capable of processing extremely large codebases, documents and extended conversations within a single session. It can accept text, images and video and generate text, while supporting tool calling and structured outputs. Access is being offered at zero cost for both input and output tokens during the preview period.
The feature attracting the greatest scrutiny, however, is not its coding ability but its treatment of developer data. The model’s notice states that prompts and completions are retained by the provider, while specifying that they are not used for training. It adds that other uses are governed by OpenRouter’s Stealth Model Terms. The operator receiving and retaining that material remains unnamed.
That creates an unusual arrangement for programmers evaluating the model on proprietary repositories, internal documentation or commercially sensitive engineering work. Developers can know that their prompts are retained, but they cannot identify the company controlling the infrastructure to which those prompts are ultimately sent.
The language also sits alongside broader terms for OpenRouter’s Stealth Program that contemplate substantially wider treatment of user content. Those terms say stealth providers generally participate anonymously while models are made available free for limited periods, with user content potentially collected and shared for model training and improvement. They also provide for content submitted to stealth models to be licensed for training, evaluation and improvement.
Ox Alpha’s individual notice therefore appears to provide a specific protection against training on its prompts and completions, while leaving retention intact and referring users back to the wider stealth agreement for other matters. The distinction is significant: retention and training are separate practices, and a promise not to train on submitted material does not mean the material is deleted immediately after processing.
OpenRouter’s general terms also warn customers that individual models may store or train on inputs under their own conditions and place responsibility on users to determine whether a model’s terms satisfy their privacy, security and compliance requirements. The service makes no guarantee regarding a provider’s data-handling, retention or security practices.
Despite those concerns, Ox Alpha has quickly attracted attention among developers because of its combination of scale and zero-cost access. Its 131,072-token maximum output allowance is unusually large, making the model potentially useful for long software-engineering tasks that require repeated reasoning across large repositories rather than isolated code generation.
A publicly documented run on the 113-task DeepSWE software-engineering benchmark reported that Ox Alpha solved 66 tasks, equivalent to 58.4 per cent. The result offers an early indication of strong agentic coding ability, although it should not be treated as an official benchmark from the model’s undisclosed developer.
The model has also generated speculation about who created it. Developers and technology observers have suggested connections ranging from Chinese AI laboratories to major Western technology companies, but no organisation had publicly claimed ownership by August 24. Similarity in behaviour to other experimental models has fuelled theories involving Z. ai and other developers, while alternative theories have pointed elsewhere. None has been substantiated.
Interest has been amplified by endorsements from technology executives and by the extraordinary volume of experimentation encouraged by free access. The model is positioned specifically for long-horizon software engineering, where agents may inspect files, execute tools, revise code and continue working through complex tasks over extended sessions.
That same use case increases the importance of its retention policy. Coding prompts can contain source code, configuration files, API details, internal architectural information and fragments of business logic. OpenRouter’s Stealth Program terms expressly caution users against supplying sensitive categories of information and make clear that material sent to a stealth model reaches the unnamed provider.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
