A complimentary AI model is gaining popularity among developers, yet no one is aware of the servers it operates on.

      Last week, an unidentified model named Ox Alpha was released on OpenRouter, available for free and featuring a context window of over one million tokens, garnering attention from developers. According to OpenRouter’s own listing, prompts and completions will be retained by the anonymous provider.

      This unclaimed model is undergoing testing in the industry. Ox Alpha debuted as a stealth launch from an unnamed third-party provider on OpenRouter last Thursday, offering no-cost access with a context window slightly exceeding a million tokens.

      The scale of this offering is significant. The open-source entity OpenCode noted that the model would be free for a week with nearly unlimited usage, highlighting that the provider has the capacity to handle 100 trillion tokens daily.

      Developers have given it high marks. Stripe CEO Patrick Collison has sampled it and described it as "very impressive," indicating its potential applications in coding, long-horizon agent tasks, and production environments.

      Speculation regarding the provider has been inconclusive. The leading hypothesis points to Z.ai, which had previously tested GLM-5 anonymously under a different name. However, an alternative analysis of its tokenizer aligns it more closely with Microsoft’s MAI family, in a market that has already been changed by free models from Chinese developers.

      By the weekend, confidence in any theory had waned. AI analyst Andrew Curran remarked that people appeared “less sure of anything” compared to the previous night.

      The more significant issue is not the identity of the provider, but the terms of use. OpenRouter’s listing indicates that prompts and completions will be "retained by the provider and are not used for training."

      Consider this with a work project in mind: whatever you send goes to a company that hasn’t disclosed its identity, and it retains that information.

      For European businesses, this is more than just a mystery; it's a significant barrier. Data protection regulations require a contract with a named processor and a review of data destinations, neither of which can occur when the opposing party remains anonymous.

      The timing amplifies the issue. The transparency obligations of the AI Act came into force on August 2, with penalties reaching up to €15 million or 3% of global turnover, in a framework that necessitates knowledge of which provider is accountable for what.

      None of this implies that the model is flawed. The release of open-weight models has outpaced the resolution of safety gaps, and a stealth launch is a valid strategy to evaluate a model before its official announcement.

      However, "free" does come with a cost: information. Someone is investing in 100 trillion tokens of inference daily, and until their identity is revealed, a practical approach for European users would be to experiment with Ox Alpha while avoiding sensitive information.