Mystery deepens over anonymous AI model Ox Alpha
The release of a free “stealth” model on OpenRouter has triggered intense debate over its origin, with speculation ranging from Chinese firm Z.ai to an unreleased Microsoft product.
A new artificial intelligence model named Ox Alpha has emerged on the platform OpenRouter, prompting a wave of speculation regarding its true developer. Released on Thursday, 20 August 2026, the model is available for free and is officially characterised as a “reasoning model designed for coding, sustained agentic work, and production workload.”
The OpenRouter listing describes the release as a “stealth model,” noting that it was developed and operated by a third-party provider who has chosen to remain anonymous during this preview period. This lack of attribution has left investors and industry observers uncertain about the strategic positioning of the new tool.
Stripe CEO Patrick Collison, whose company is currently in the process of acquiring OpenRouter, publicly praised the model on X, describing it as “very impressive.” While the endorsement adds weight to the model’s perceived capabilities, the ongoing business relationship between Stripe and OpenRouter provides a layer of commercial context to the release.
Initial online debates focused heavily on the possibility that the model was created by Chinese company Z.ai, known for its GLM series of AI models. However, AI analyst Andrew Curran noted on Friday that this specific line of speculation has lost momentum, with observers now appearing less certain of the origin.
As the mystery persists, alternative theories have gained traction. An update from Wccftech suggests that Ox Alpha could be an unreleased version of Microsoft’s MAI model series. On social media platforms such as Reddit, opinions remain sharply divided, with some users expressing high confidence in a Chinese origin while others argue the model cannot be from China.
The true identity of the developer behind Ox Alpha remains unconfirmed, leaving the broader market to weigh the model’s performance against the uncertainty of its provenance.
