Unmask a Stealth LLM: Fingerprint Any Model API

On 20 August 2026 a model called stealth/ox-alpha showed up on OpenRouter with no lab name, no logo, no press release, a 1,048,576-token context window, and a free preview window that ran about a week. Within 48 hours it was in production traffic at Nous Research's Hermes agent and inside the Zed editor. Nobody could say who made it.

The interesting part was not the guessing. It was the method the community converged on within a day: stop asking the model who it is, and start measuring the plumbing around it. Independent researcher Ben Davis put it at Zhipu's unreleased GLM-5.x series with 99% confidence, leaning mostly on video-encoder token consumption and tokenizer alignment. Separately, unclecode (author of Crawl4AI) collected the scattered one-off tricks people were posting into modelprint, a single-page tool with an approved registry of 15 probes. Its day-one ranking put z-ai/glm-5.3 at 6 of 9 probes matched, with all four normalized tokenizer counts hitting exactly. No other lab's best candidate cleared more than 2 of 4.