Nobody knows who built AI coding model Ox Alpha or where the code goes
SiliconANGLE Duncan Riley ● Covered by 3 sources
A mystery coding model called Ox Alpha is free on OpenRouter, and no one’s claiming it. It’s already chewing through enterprise code while its builder — and where the data goes — stay unclear.
Based on reporting by SiliconANGLE, Duncan Riley — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
A frontier-style coding model has been sitting on OpenRouter since Aug. 20 with a simple label: Ox Alpha. No company has stepped up to say it built the thing. The price is hard to argue with — nothing for input or output tokens — which is probably why developers started poking at it right away.
The model isn’t small-time. Its context window stretches to 1,048,576 tokens, it can produce up to 131,072 tokens, and it accepts text, images and video. Audio gets rejected. OpenRouter describes it as a reasoning system meant for long-horizon software engineering and sustained agentic work, which is a very specific way of saying: hand it big jobs and see how long it can keep its head straight.
Early benchmark chatter gave it traction. Ben Davis, who posts as @davis7, ran Ox Alpha on 10 DeepSWE tasks and got it above 80%, ahead of Claude Fable 5 at 65% and GPT-5.6-sol at 52%. Davis also said the tiny sample size could swing the result, and when he later tested the full DeepSWE set, Ox Alpha landed roughly level with GPT-5.6-sol mid rather than dramatically above it. No public leaderboard has published a score.
The attribution hunt has been more interesting than the score chase. A developer known as unclecode, who built the Crawl4AI crawler, made a fingerprinting tool called modelprint that compares anonymous API endpoints against known models. On that test, Ox Alpha matched GLM-5.3 on six of nine infrastructure probes, including all four normalized tokenizer counts. That still doesn’t prove identity, only shared infrastructure. And the guess is not settled: Xiaomi’s MiMo team has come up in some discussions too, while one tokenizer reading points toward cl100k_base, an encoding tied to OpenAI and a strange fit for a Chinese model.
The terms are where this gets sharp. OpenRouter says prompts and completions are retained by the provider and not used for training, while broader anonymous-preview terms on the site cover training, evaluation and improvement. OpenCode’s route says zero retention, but it doesn’t name the provider. If the Z.ai theory is right, that matters even more, because Zhipu AI — Z.ai’s former name — was added to the U.S. Commerce Department’s Entity List on Jan. 16, 2025. Bloomberg reported Sunday that coding tools including Claude Code have already pushed billions of tokens through Ox Alpha since Thursday. So the mystery model is not sitting idle. It’s already in the workflow.
My take — AI-written commentary, not fact-checked reporting
Anonymous models are fun until they start handling real company code and nobody can say where the prompts land. Open access with fuzzy terms is basically the AI industry’s favorite way to ask for trust without earning it. If a provider won’t even attach a name, the right default is suspicion, not swagger.
Read more about this at: SiliconANGLE