A new model named Ox Alpha, released quietly via OpenRouter on August 20, 2026, is sparking debate in AI circles over who actually built it. The model is labeled “stealth,” allowing developer access through OpenRouter but hides its creator. It’s explicitly marketed for heavy-duty work—coding, long agentic workflows, and production usage.([glm5.app](https://glm5.app/blog/ox-alpha-openrouter?utm_source=openai))
What We Do Know: Specs and Features
Ox Alpha appears with the model ID stealth/ox-alpha. It supports multimodal input—text, images, and video—and outputs text. Most notably, it has a massive 1,048,576-token context window and can generate up to 131,072 tokens in a single output during its preview stage.([glm5.app](https://glm5.app/blog/ox-alpha-openrouter?utm_source=openai))
The model is free for both prompts and completions during this preview period. One standout detail: reasoning is mandatory. Users can’t disable it—they can only adjust its intensity among “low,” “high,” or “max,” with “max” set as the default. Tool calling and structured or JSON-style outputs are supported.([glm5.app](https://glm5.app/blog/ox-alpha-openrouter?utm_source=openai))
What Remains Hidden
Despite its strong technical appearance, Ox Alpha’s origin remains unclear. OpenRouter lists it under a “Stealth” provider who chose anonymity for the preview. There is no information at this time about who developed it, how many parameters it has, or the precise architecture behind it.([glm5.app](https://glm5.app/blog/ox-alpha-openrouter?utm_source=openai))
There are no official benchmark scores published. Community-driven evaluations so far include DeepSWE—a coding-agent benchmark—where Ox Alpha passed about 80% of a small sample test of 10 tasks. But experts caution these results remain anecdotal and limited.([techtimes.com](https://www.techtimes.com/articles/325244/20260823/coding-model-ox-alpha-retains-every-prompt-you-cannot-name-company-holding-them.htm?utm_source=openai))
Who Might Be Behind It?
Multiple developers and forensics reports point toward Zhipu AI / Z.ai as the most likely source. Tokenizer fingerprints, error-message similarities, and video encoder tests align Ox Alpha with the GLM-5.3 lineage. Still, it’s unclear whether Ox Alpha is simply public GLM-5.3, a hidden checkpoint, or a broader variant.([theclarity.today](https://theclarity.today/story/ox-alpha-debuts-anonymously-on-openrouter-2a29eeca?utm_source=openai))
Other possibilities floated include Microsoft’s MAI branch, but those hold less weight in current discussions. Conversations across GitHub, Reddit, and private forensics reinforce the GLM connection, though no owner has officially confirmed or denied the attribution.([techcrunch.com](https://techcrunch.com/2026/08/23/whos-behind-the-new-stealth-model-ox-alpha/))
Operational and Privacy Considerations
Ox Alpha is in a preview state—it’s plausibly coming with rate limits, free access, and anonymized terms that could change. All prompts and responses are logged by the model’s provider. Though the listing states those logs are not used for training, the retention policy remains relevant for teams using sensitive data.([glm5.app](https://glm5.app/blog/ox-alpha-openrouter?utm_source=openai))
Reliability considerations include a single-provider route via OpenRouter; this means if “Stealth” suffers downtime, any system using Ox Alpha could be blocked. Additionally, a preview-free pricing regime may give way to paid tiers or other restrictions once the preview ends.([glm5.app](https://glm5.app/blog/ox-alpha-openrouter?utm_source=openai))
For developers eager to try it, use cases that play to Ox Alpha’s strengths involve long-term coding projects, multi-file refactors, workflows combining visual evidence and code, and agentic sequences rather than simple Q&A. But until benchmark data is published, results should be validated against real-world tasks rather than hype.([glm5.app](https://glm5.app/blog/how-to-use-ox-alpha?utm_source=openai))
Ox Alpha represents a fresh flashpoint in AI’s steady march toward larger context windows, reasoning-enforced models, and anonymous releases. If it truly comes from the GLM-5.3 lineage, this could mark a shift in how labs deploy high-capability models: more stealth launches, less upfront marketing, and heavier reliance on community for events like attribution. For users, the experiment seems promising — but trust must be earned. Watch closely for vendor disclosures, real benchmark data, and whether Ox Alpha behaves as well at scale as on paper.
