The detective work is worth your time. Four independent evidence classes were used to fingerprint Ox Alpha: a Java stack trace pointing to Z.ai infrastructure, error code 1214 matching their API, 30-of-30 tokenizer matches to GLM-5.3, and 95-of-95 exact vocabulary hits. This is the fifth stealth model drop on OpenRouter in six months. All four previous ones were eventually claimed. Stealth-model forensics is becoming a community sport with actual methodology — discriminating probe strings, chat-template grammar analysis, the works.
Meanwhile, 221,000 developers sent their code to an entity they literally cannot name. Proprietary code, presumably. The anonymous provider claims capacity of 100 trillion tokens per day, which is hyperscaler-scale spend. Somebody is bankrolling this and choosing not to say who.
The security angle makes it weirder. Z.ai says GLM-5.3's cybersecurity capabilities grew faster than expected during training — the model found over a thousand real bugs in widely used software. Their response: hold back the public weights for about two weeks of extra safety review before open release. Compare that to Anthropic, which discovered similar emergent security capabilities in Claude Mythos back in April and locked the entire model under ASL-4 restrictions.
Both labs hit the same surprise during training. Anthropic locked the model away. Z.ai appears to have already put theirs in front of a quarter million users under a pseudonym while finishing the safety review.