이미지: openrouter.ai 화면 갈무리
Summary
- An unidentified company's reasoning model called 'Ox Alpha' has been listed for free on OpenRouter
- It supports a 1,048,576-token context window, up to 131,072 tokens of output, and text, image, and video input
- OpenRouter's page states explicitly that it is not the developer, only a relay for requests
- 모델명
- Ox Alpha (stealth/ox-alpha)
- 공개일
- 2026년 8월 20일
- 개발·운영 주체
- 정체를 밝히지 않은 제3자 업체, 오픈라우터는 중계만 담당
- 컨텍스트 윈도우
- 1,048,576토큰(약 1M), 최대 출력 131,072토큰
- 입출력
- 텍스트·이미지·비디오 입력 → 텍스트 출력, 도구 호출(tool calling) 지원
- 가격
- 무료(프롬프트·완성 토큰 과금 없음)
- 성능(P50)
- 지연시간 5.57초, 처리량 24 tokens/s, 3일 가동률 99.99%
A model with no name behind it, and no maker disclosed
On August 20, a new model called 'Ox Alpha' appeared on OpenRouter. It's introduced as a reasoning model meant for coding, long-running agent tasks, and production use, but there's no information on who actually built it. The Ox Alpha model page simply states: "This stealth model is developed and hosted by a third-party model provider," without naming the company.
OpenRouter is a service that bundles models from multiple AI companies behind a single API. Here too, OpenRouter only connects users to Ox Alpha through an OpenAI-compatible API, stating flatly on the page that it is "not the developer, owner, or provider of this model." It notes that prompts and completions are stored by the provider and are not used for training, but adds that the provider's identity will remain anonymous for the duration of the preview period.
What the model can do
According to the description, Ox Alpha is built for long-horizon software engineering tasks, complex reasoning, and workflows that combine text and visual information. It accepts text, image, and video input and produces text output. Its context window is 1,048,576 tokens — large enough to load and reference a codebase or set of documents spanning several books' worth of content in a single pass. Maximum output is 131,072 tokens. It supports the tools and tool_choice parameters for function calling and can output JSON, though the page notes it does not support enforced JSON schemas.
Specs and performance metrics
| Metric | Value |
|---|---|
| Context window | 1,048,576 tokens (~1M) |
| Max output | 131,072 tokens |
| Latency (P50) | 5.57 sec |
| Throughput (P50) | 24 tokens/s |
| 3-day uptime | 99.99% |
| 3-day availability | 99.14% |
| Price | Free |
The page notes that since there is only one provider, OpenRouter doesn't need to handle any additional routing logic. Real-time metrics such as uptime and latency are continuously updated on the OpenRouter status page, which currently shows 99.99% uptime and 99.14% availability — a measure of successful response rates — over the past three days.
Why hide the identity?
Releasing a new model anonymously ahead of an official launch to gauge reaction is not unusual in the AI industry. Rather than relying on benchmark scores or marketing copy, providers test the model against real user traffic first, and if reception is strong, relaunch it under its official name. OpenRouter's phrasing that the provider "chose to remain anonymous during the preview period" fits this pattern. Still, there's nothing in the available information to indicate which company might be behind it.
How to try it
Ox Alpha can be accessed directly through OpenRouter's website or its API.
- Visit the Ox Alpha model page and click the 'Try this model' button to open the playground.
- Enter a prompt in the playground to get a response immediately, with no additional setup required.
- To connect via API, point the base URL of your existing OpenAI-compatible SDK to OpenRouter's endpoint and set the model slug to
stealth/ox-alpha. - Both prompt and completion tokens are free, so no payment information needs to be entered.
No country or plan restrictions are listed for use. It appears accessible with just an OpenRouter account. The 1M-token context could be tested by loading an entire large codebase for refactoring, or by referencing design documents that include images to generate frontend code. Since it supports tool calling, it could also be run as an agent by connecting external functions such as web search or file manipulation.
Editor's take
This isn't the first time a stealth model has shown up on OpenRouter. The pattern repeats for a simple reason: benchmark scores are never fully free of gaming controversies, but real failures that users hit while actually coding can't be faked. By releasing a model without a company name attached, users are forced to judge it purely on output, free of brand perception. It's no accident that Ox Alpha targets coding and long-running agent work head-on — the two arenas where competition in the industry is currently fiercest. Drop it anonymously, see how it lands, and quietly pull it if the reaction is bad.
Putting a model like this to real use always runs into the same dilemma. The 1M-token context and free pricing are appealing, but the fact that prompts and outputs are stored by an unidentified company is not something to brush aside. Even with an explicit statement that the data isn't used for training, feeding prompts containing internal codebases or customer data into a model like this is a call that needs to separate personal experimentation from company work. It carries little risk for an individual testing out a new capability, but production code or sensitive documents should wait until after an official launch.
There's a good chance the model's identity will surface within the next few weeks. Stealth models are typically relaunched under an official name once performance is validated, or quietly disappear if reception falls flat. With throughput of 24 tokens/s and a first-response latency of 5.57 seconds — numbers that fall short of top-tier — whether this model eventually launches under a real brand or simply fades away will depend on user reaction over the coming weeks.
Comments