Ox Alpha, Stealth AI: Uncertain Origins and Free Access

Ox Alpha, a reasoning model that quietly arrived on OpenRouter, is free and presented as tailored for coding and agentic tasks. Its origin is the subject of contradictory hypotheses, ranging from Z.ai (GLM-5) to a connection with Microsoft's MAI family, with no definitive proof. The OpenCode agent claims a daily capacity of 100 trillion tokens, and Stripe CEO Patrick Collison finds the tool very impressive. This phenomenon occurs as several Chinese labs claim performance levels close to American systems at a lower cost.

Contested Origin Between Chinese Lead and Microsoft Connection

Several public analyses contradict each other regarding the origin of Ox Alpha. Wccftech suggested that Z.ai, the creator of GLM-5, might be behind the model, recalling that Z.ai had previously tested GLM-5 anonymously under the name Pony Alpha and noting that developers have observed similarities between the behavior and responses of Ox Alpha's tokenizer and those of GLM-5. The same publication later highlighted another analysis that would instead link Ox Alpha's tokenizer to Microsoft's MAI family. Beyond these hypotheses, rumors more generally mention either a Chinese AI lab or a connection with Microsoft. On Saturday, analyst Andrew Curran indicated on X that while the GLM lead prevailed on Friday night, uncertainty had risen again by the following morning. At this stage, the available evidence does not allow for a conclusion.

On OpenRouter, a Stealthy Free Model Tailored for Coding

Ox Alpha appeared on Thursday on OpenRouter, presented as a stealth model provided by an anonymous third party. OpenRouter describes it as a reasoning model designed for coding, supported agentic work, and production loads, suitable for long-term software engineering, complex reasoning, and workflows combining text and visual context. The model is accessible for free. OpenCode clarified on X that this free access would last for a week with nearly unlimited usage and claimed that the provider could handle 100 trillion tokens per day, which is about 100 times the monthly volume of AI tokens reported by Visa. In terms of initial feedback, Patrick Collison, CEO of Stripe, tested Ox Alpha and found it very impressive.

Increased Competition: Zhipu, DeepSeek, Moonshot AI Lurking

The emergence of Ox Alpha occurs in a context where several Chinese labs, including Zhipu, DeepSeek, and Moonshot AI, challenge their American counterparts. According to descriptions, their models exhibit performance levels comparable to those of American benchmark systems, while being offered at a significantly lower cost and providing extensive access to code. In July, Moonshot introduced Kimi K3, a model with 2.8 trillion parameters, aimed at coding, reasoning, and various agentic tasks, which quickly gained attention in Silicon Valley due to its results and attractive pricing.