A coding model with a context window of roughly one million tokens appeared on OpenRouter with no company name, no logo, and no press release. It is called Ox Alpha, and the only provider label attached to it is Stealth. Within days the developer community began working out where it came from, and by August 22 the accumulated evidence pointed at the serving infrastructure of China's Zhipu AI, known internationally as Z.ai. Neither confirmation nor denial has been issued.

A High-Performance Model That Arrived Unnamed

Ox Alpha was spotted on August 20, 2026. It sat in OpenRouter's model catalog as stealth/ox-alpha with nothing to identify its developer. OpenCode, the open-source coding agent, said it would be free for about one week, meaning through roughly August 27. The anonymous provider claims capacity of 100 trillion tokens per day.

The first thing developers noticed was the size of the context window: 1,048,576 tokens, or about one million. Given that most production-grade models are designed for somewhere between 8,000 and 200,000 tokens, this is less an incremental improvement than a change in what the tool can be used for. A sizable codebase can be loaded into a single prompt and reasoned over as a whole.

Inputs include text, images, and video, and the model supports function calling and structured JSON output. Throughput has held at around 29 tokens per second, a reasonable figure for a model handling this much context.

Four Classes of Evidence Assembled by Developers

The identification effort proceeded through measurement rather than impression.

The first tool was tokenizer fingerprinting. A tokenizer splits text into the numeric units a model processes, and each model family has its own vocabulary and segmentation habits, so token counts for identical inputs become a usable signal. When a researcher ran 30 probe strings covering 14 writing systems, emoji, code, and SQL, every one matched GLM-5.3's tokenizer. Each request also came back exactly 75 tokens higher, which indicates an invisible system prompt or routing wrapper prepended to every query rather than a different architecture.

The second was video encoder behavior. Ox Alpha accepts video input, and three separate design choices lined up simultaneously with Zhipu's multimodal model GLM-5V-Turbo: how many frames are sampled per second, how token cost scales with duration (roughly 147 tokens per second), and how token cost changes with frame resolution. One number matching could be coincidence. Three independent choices matching at once is much harder to explain.

The third was decisive. A deliberately malformed request, with an invalid value in a parameter, caused the server to return a Java error with a stack trace that exposed the internal class name com.wd.paas.api.domain.v4.chat.ChatCompletionRequest. That package path corresponds to the structure of Zhipu's published API.

The fourth was the error-code dialect. Given an invalid role field, Ox Alpha returns code 1214 with the message Incorrect role information. That is the same format Z.ai-hosted GLM models return, while a different operator hosting the same GLM-5.2 weights returns a different error format. In other words, this evidence identifies the operator serving the model, not merely the weights being served.

Researchers involved in the work put their confidence at the operator layer at 0.98, or 99 percent. Even so, neither Zhipu AI nor OpenRouter has said anything on the record.

Treat the Benchmark Numbers as Preliminary

On performance, a figure making the rounds is 8 of 10 tasks passed on DeepSWE, which evaluates models on real software engineering problems, for an 80 percent pass rate. Claude Fable 5 reportedly scored 65 percent and GPT-5.6-Sol 52 percent on the same tasks.

These are not audited leaderboard results. They come from a volunteer run of just 10 tasks, and with only 10 tasks a single outcome moves the pass rate by 10 points. It is also not a benchmark the major labs quote, so it is best read as an early signal of direction. Attention accelerated sharply after Patrick Collison, CEO of Stripe, which agreed to acquire OpenRouter on August 19, spoke favorably of the model, which makes it easy for the numbers alone to take on a life of their own.

Stealth Launches Are Already a Standard Playbook

Ox Alpha is the fifth stealth model to appear on OpenRouter in six months. The previous four were Z.ai's own GLM-5, Xiaomi's MiMo-V2-Pro, which showed up under the names Hunter Alpha and Healer Alpha, Ant Group's Lingxi Ling-2.6-flash, and Meituan's LongCat-2.0. In every case the developer stepped forward once the preview period ended.

The advantages of releasing anonymously are clear. When a lab's name is attached, the evaluator's expectations get mixed into the result. Without a name, the model is judged on its outputs. The lab, in turn, collects a large volume of stress-test data under conditions close to real workloads, plus a user base already comfortable with the product at the moment of the official announcement. Serving 100 trillion tokens a day of free inference is not cheap, but the math works if some share of those users converts to paid use.

The Line on What You Should Send Is Blurry

In practice, the thing to watch is not performance but data handling.

The OpenRouter page for Ox Alpha states that prompts and completions are retained but are not used for training by the provider. The terms that govern stealth models across the platform, however, grant OpenRouter and the anonymous provider the right to use user content for training, evaluation, and improvement, and instruct anyone unwilling to accept that to refrain from using stealth models. They also specify that personal data contained in inputs will be sent to the provider. How these two are reconciled cannot be determined from the published materials.

Access through OpenCode is described differently, with zero retention and no training use. The same model therefore carries different stated policies depending on the route. If the content is sensitive, that difference is not something to wave away.

If the provider really is Z.ai, another set of considerations applies. Z.ai is the international name of Beijing-based Zhipu AI, which listed on the Hong Kong Stock Exchange in January 2026. Organizations headquartered in China fall under Article 7 of the National Intelligence Law, and the company was added to the US Commerce Department's Entity List in January 2025. In May 2026, US House committees opened an inquiry into security concerns around Chinese AI models in critical infrastructure, with the company among those named. These are structural legal conditions rather than accounts of anything that has happened, but they are worth weighing before routing a codebase through the service.

Summary

Ox Alpha was released anonymously and free of charge as a coding model with a one-million-token context window and support for image and video input. Four lines of investigation, covering the tokenizer, the video encoder, a stack trace, and error codes, all point to Zhipu AI's GLM family running on Z.ai infrastructure. Official confirmation is still absent, and the benchmark figures come from a 10-task preliminary run. The free window is said to close around August 27, so it is worth trying, but decide what to send it only after checking the policy attached to the route you are using.