Ox Alpha : The Stealth Frontier Model Free For a Week - 1M Context

Original author：AI WITH Rithesh

Original source：https://www.youtube.com/embed/zNyJRkQiZww

A frontier-class model appeared on OpenRouter on 20 August 2026 with a 1,048,576-token context window, text/image/video input, 131,072 tokens of output, and no company name attached to it. It is free $0 in, $0 out for a week, on a gateway that says it has capacity for 100 trillion tokens a day. Within twelve hours one developer had pushed 4 billion tokens through it, another was calling it possibly the best frontend model available right now, Nous Research had wired it into Hermes Agent, and four different labs were in the frame for having built it. WHAT IT CAN DO. The shape is the story: a million tokens of context holds a codebase rather than a file, 131K of output writes a whole feature in one response instead of dribbling it out over twenty turns, and video input is rare at this level. That is a model built for agents that run for hours. In practice, day-one testers reported the spinning donut one-shotted, a black hole render posted as a "masterpiece", frontend and 3D simulation work run head-to-head against Grok 4.6 on the same prompt, and one very well-known developer merging eight pull requests without reading the code because the model said they were fine. WHO BUILT IT. Nobody has claimed it. The strongest argument is precedent, not evidence: every prior stealth model on this gateway turned out to belong to a Chinese lab, and the last one ran anonymous for two months before its lab confirmed it and open-sourced the weights the next day. Theory one, currently loudest, is a Z.ai/GLM model most likely the larger teacher behind GLM-5.3 rather than GLM-5.3 itself. Theory two is Xiaomi's MiMo team, on a shared tokenizer fingerprint — except the same broken token also trips Qwen and GLM, so it proves "a Chinese lab", not which one. Theory three says the country is wrong entirely and floats Grok-4.7 or Gemini 3.8 Flash. And the most-liked take of the launch calls it frontier, says it is good in the ways Chinese labs are usually weak, and half-jokes that it could be a stolen Claude checkpoint. WHAT NOBODY KNOWS. It is on no independent leaderboard — not LMArena, not Artificial Analysis. The circulating numbers come from one day-one automated harness (97% maths, 95% coding, 96% reasoning — plus three separate 100%s on hallucinations, ethics and general knowledge, which says more about the test than the model) and one person's 10-task DeepSWE run at 80%. The row nobody quoted is instruction following at 77%, the lowest of the eight and the one that decides whether a long agent task goes the way you asked. Also covered: testers who cannot agree whether it is fast or slow, the missing cybersecurity guardrails, and the conflict between the gateway's "Zero Data Retention" and the model page's "prompts are retained by the provider". Every attribution here is quoted as somebody's read, at their own stated confidence. No lab has confirmed anything. Everything in this video was published within 36 hours of upload if the model gets claimed after this goes live, treat the reveal as the correction. SOURCES OpenRouter model page — https://openrouter.ai/stealth/ox-alpha OpenRouter launch post — https://x.com/OpenRouter/status/2090544970923184269 OpenCode free-week post — https://x.com/opencode/status/2090544355824038300 Tokenizer fingerprint — https://x.com/AiBattle_/status/2090555139899654456 Grok / Gemini theory — https://x.com/ai_for_success/status/2090662996003692815 Frontend head-to-head vs Grok 4.6 — https://x.com/SPAC89/status/2090587270457840034 One-shot donut — https://x.com/yashcodes2007/status/2090668131287220719 Black hole render — https://x.com/bisheshab/status/2090562869276618807 8 PRs merged unread — https://x.com/theo/status/2090669658483691539 GLM read — https://x.com/scaling01/status/2090662468833976582 "Frontier. wtf." — https://x.com/teortaxesTex/status/2090631974209626297 GLM 5.x + vision read — https://x.com/teortaxesTex/status/2090680807400620087 4B tokens in 4 hours — https://x.com/mehulmpt/status/2090621816167596486 DeepSWE 10-task run — https://x.com/NFT_Chen/status/2090684130153611637 Hermes Agent — https://x.com/Teknium/status/2090674052058984513 Day-one harness — https://benchable.ai/models/stealth/ox-alpha Stealth-model background — https://www.orcarouter.ai/blog/ox-alpha-stealth-model-what-we-know Owl Alpha reveal — https://dayverse.id/en/articles/true-identity-of-owl-alpha-on-openrouter-revealed/ No lab has claimed Ox Alpha. Every attribution above is somebody's read.