Out of nowhere, an unheralded AI model called Ox Alpha popped up as a free preview on OpenRouter—and instantly set the developer community on fire.
Arriving as a full multimodal agent capable of processing text, images, and video, Ox Alpha started pulling down monster scores on long-context coding benchmarks. Naturally, everyone immediately wanted to know one thing: who actually built this?
The DeepMind Teases vs. The Digital Sleuths
The hype machine hit overdrive when social media posts from Google DeepMind insiders began making the rounds. DeepMind researcher Vamsi Batchu dropped a cryptic “Google is back. Trust the process,” while Jonathan Ouyang posted a simple “It’s Gemini time :)”.
That was all it took. Within hours, threads across X and Reddit declared Ox Alpha a leaked, unreleased Gemini variant—making it a huge talking point across mainstream Google AI news cycles.
But while the internet was celebrating a secret Google launch, digital sleuths were busy digging into the backend. Community researchers ran forensic probes against the model’s responses, and the evidence quickly pointed away from Mountain View and straight toward China’s Zhipu AI:
What Sleuths TestedWhat They FoundWhat It Pointed ToTokenizer Probe30 out of 30 exact token matches across code and math.Identical to Zhipu’s GLM-5.3 tokenizer.Server Error CodesTriggered Java stack traces returning raw error code 1214.Matches Zhipu’s internal Z.AI infrastructure.Video ProcessingExact frame-to-token count parity under stress tests.Mirrors Zhipu’s GLM-5V-Turbo encoder.
Neither Zhipu AI nor OpenRouter has officially claimed the model yet, leaving it in a bizarre state of anonymous limbo.
What the Ox Alpha Drama Tells Us About AI Rumors
If the Ox Alpha saga proves anything, it’s that the AI ecosystem moves on pure vibe velocity right now. A pair of well-timed tweets from engineers can overshadow hard technical data for days before anyone checks the error logs.
At the same time, it shows how sharp the community has become at holding the hype accountable. Tokenizer matching, stack trace analysis, and dialect probes are now standard community toolkit items whenever a mystery model surfaces alongside breaking Google AI updates.
Ox Alpha remains live for developers looking for a fast, free multimodal coder. But if you were hoping you were testing Google’s next flagship engine in secret, the data says you’ll probably have to keep waiting.
Add NPowerUser as a preferred source on Google News