Around August 20, 2026, Ox Alpha appeared on OpenRouter under the name stealth/ox-alpha — no company announcement, no technical white paper. Independently confirmed specs: a 1.05M-token context window, 130K max output, text/image/video input — the first natively omnimodal model in the Stealth anonymous series. It launched with a free week of usage ($0 per million tokens), a zero-data-retention policy, a knowledge cutoff extending to November 2025, and a cache hit rate of about 77.9%. See our original report: Ox Alpha quietly lands on OpenRouter.
Early community tests showed Ox Alpha beating GPT-5.6 Sol and Fable 5 on coding and agentic tasks; its 1M context matches Gemini 3.7 Flash while adding video input and being free for now — versus DeepSeek V4 Flash's 128K context. Worth watching, but apply two discounts: an anonymous model has no official benchmark report, and community tests vary wildly in method and sample size; and free-period serving configs may not match the eventual production release.
Our standing advice on viral scores matches the conclusion of our Qwen 3.8 27B benchmark-debate guide: unsigned scores with opaque methodology should be treated as pending independent verification — small-scale tests on your own real tasks remain the most reliable check.
No signature, but the technical fingerprints are dense: the tokenizer matches Zhipu's GLM family on all 11 measured indicators; in a digit-probe test Ox Alpha consumed 29 tokens — an exact GLM match (DeepSeek uses 98); its deployment configuration closely mirrors GLM-5.3, and Ox Alpha adopted GLM-5.3's new settings just two days after they shipped. Zhipu founder Tang Jie had also teased an upcoming multimodal mode three months earlier. Previous Stealth anonymous models were eventually revealed as GLM-5 or MiMo releases — the "anonymous warm-up to collect high-concurrency data" playbook has precedent. The most likely reading: Ox Alpha is a stealth test of Zhipu's multimodal GLM.
Running parallel to the sober fingerprint analysis was a classic internet frenzy: in the early hours of August 21, an anonymous GitHub Gist showed Ox Alpha solving olympiad-level math; two prominent AI-safety researchers posted cryptic tweets ("we are not just facing a new model") and quickly deleted them; Reddit and Hacker News spent 72 hours speculating — an OpenAI shadow experiment? A reverse-engineered artifact? Performance art? OpenAI and Google declined to comment. See our report: Mysterious model Ox Alpha ignites online speculation.
The frenzy itself is the lesson: when a technology is too strong to verify, imagination fills every gap — keeping the line between fact and speculation matters more than picking a side.
If your use case involves whole-repository code understanding or multimodal input, the free quota is worth a try: the 1M context lets you feed an entire codebase for refactoring without RAG chunking; video input can help diagnose front-end interaction bugs; and the API is OpenAI-SDK compatible, so it drops into Cursor, Aider and similar tools quickly.
Three cautions: keep a known model like GPT-5.6 Sol as a baseline for comparison; "zero data retention" is a platform claim — stay careful with sensitive data; and an anonymous test model can vanish or be repriced at any time, so do not bet a production pipeline on it — evaluate long-term pricing and SLAs after an official release.
Ox Alpha is a large model that appeared anonymously on OpenRouter as stealth/ox-alpha around August 20, 2026: a 1.05M-token context window, 130K max output, text/image/video input, free for a limited time at launch with a zero-data-retention policy. No developer has claimed it.
Not officially confirmed, but the technical fingerprints point strongly to Zhipu's GLM family: the tokenizer matches GLM on all 11 measured indicators, a digit-probe test consumed 29 tokens (an exact GLM match), and the deployment config closely mirrors GLM-5.3. Past Stealth anonymous models were all eventually revealed as GLM-5 or MiMo releases — the most likely explanation is a stealth warm-up test of Zhipu's multimodal GLM.
Early community tests showed it beating GPT-5.6 Sol and Fable 5 on coding and agentic tasks, but these are unofficial tests with inconsistent methodology, and free-period serving may differ from a production release. Treat the scores as pending independent verification: test on your own real tasks and keep a known model as a baseline.
At launch it offered a free week via OpenRouter and OpenCode Zen ($0 per million tokens). Availability and pricing of an anonymous test model can change at any time — check the current OpenRouter listing before relying on it, and do not build production pipelines on top of it.
