Ox Alpha (stealth/ox-alpha) is an anonymous AI model that appeared on OpenRouter on August 20, 2026 under the “Stealth” provider. Nobody has claimed it. It is free to use during a limited preview, accepts up to 1,048,576 tokens of context, and is positioned for coding, long-running agentic work, and production workloads. This page collects everything that is actually verified about it — and clearly separates that from speculation.
Ox Alpha at a glance
| Model ID | stealth/ox-alpha |
| Provider | Anonymous (“Stealth” on OpenRouter). OpenRouter routes the model; it did not build it. |
| Listed | August 20, 2026 |
| Price | $0 input / $0 output during the preview |
| Context window | 1,048,576 tokens |
| Max output | 131,072 tokens |
| Input | Text, image, video (audio is rejected) |
| Output | Text |
| Capabilities | Reasoning, tool / function calling, structured JSON output, agentic and coding focus |
| Open weights | No |
| Official benchmarks | None published |
| Free window | Expected to end around August 27, 2026 (not guaranteed) |
What “stealth model” means
OpenRouter periodically lists unreleased models from AI labs under a generic “Stealth” provider with a codename ending in Alpha. The lab gathers real-world usage before launch; in return, the model is free for a short window. Every previous stealth model was eventually revealed under its official name — see the live tracker and reveal history. Ox Alpha follows the same pattern.
What the 1M context window gets you
1,048,576 tokens is roughly the size of several full-length books, or a large codebase, in a single request. In practice that means you can paste entire repositories, long specification documents, or hours of meeting transcripts and ask questions across all of it without chunking or retrieval pipelines. Combined with a 131,072-token output limit, Ox Alpha can also produce very long responses — full refactors, multi-file patches, or long reports — in one turn.
Multimodal input
The listing accepts text, images and video as input and returns text. Audio input is rejected. Screenshots, diagrams, UI mockups, and short clips can be sent alongside a prompt through the OpenRouter API.
Where you can use it
- OpenRouter — the primary listing. How to use Ox Alpha for free walks through it in five minutes.
- Right here — chat with Ox Alpha in your browser using your own OpenRouter key. Your key and prompts never touch our server.
- Other gateways — the model is also listed via OpenCode Zen, AI/ML API, AIHubMix, ModelsLab and Routin.
Data and privacy terms
Ox Alpha runs under OpenRouter’s Stealth Program terms. Prompts are retained by the anonymous provider; the listing states they are not used for training. Because you do not know who is on the other end, treat it like any pre-release service: do not send personal data, credentials, or confidential material.
Who made it?
Officially unknown. Independent fingerprinting analyses circulated by testers point toward the GLM family, with a minority view suggesting Xiaomi’s MiMo line. Both are unconfirmed speculation until the developer announces the model. We lay out the evidence in Who made Ox Alpha? and track third-party results on the benchmarks page.
Not to be confused with
“Ox” is a common name in tech. Ox Alpha the model is unrelated to OX Security, the Ox productivity tools, Oxus.AI, or “Ox Ai Bot”.