Ox Alpha: The Anonymous AI Model With a 1M-Token Context Window

Ox Alpha (stealth/ox-alpha) was an anonymous reasoning model that appeared on OpenRouter on August 20, 2026. OpenRouter described it as a model for efficient coding, sustained agentic work, and production use, but listed the provider only as “Stealth.” The developer had chosen to remain anonymous during the preview, and no company had publicly claimed the model by the August 21 reporting cutoff. 72141

Its combination of free access, multimodal input, and an unusually large context window quickly made it a focus of developer testing. The early coding results were striking, but they were also preliminary: the available evidence supports treating Ox Alpha as a promising preview, not as a verified overall leader.

Ox Alpha’s published specifications

The model was offered at zero cost during an initial preview described as lasting about one week. Its listed context window was 1,048,576 tokens, with a maximum output of 131,072 tokens. It accepted text, images, and video, and was positioned for coding and long-running agent workflows. 2122

Specification	Reported detail
Model ID	stealth/ox-alpha
Provider	Anonymous “Stealth” provider
Release date	August 20, 2026
Preview price	Free during the preview
Context window	1,048,576 tokens
Maximum output	131,072 tokens
Input types	Text, image, and video
Intended uses	Coding, sustained agent work, and production workloads

A context window of this size can be useful for repository-scale coding tasks, long documents, and agent sessions that need to retain more working material. It does not, by itself, prove that the model will reason reliably across everything placed in that context; context capacity and practical performance are different measurements.

The early DeepSWE result was impressive—but limited

The result that generated the most attention came from an independent run of 10 DeepSWE tasks. That test reported an approximately 80% result for Ox Alpha, compared with 65% for Claude Fable 5 and 52% for GPT-5.6 Sol. 611

Those figures suggest that Ox Alpha performed strongly in that tester’s setup, particularly on software-engineering tasks involving real codebases and patches. They do not establish a definitive ranking for several reasons:

• The test covered only 10 tasks from the broader benchmark.
• It was an independent evaluation rather than an official model submission.
• The results were not verified on the official DeepSWE leaderboard.
• Small samples can be sensitive to task selection, prompting, model settings, and near-miss judgments.

Reporting explicitly noted that Ox Alpha had not been added to the official DeepSWE leaderboard. 12 The fairest conclusion is therefore that the model showed a strong early signal—not that it had conclusively surpassed every comparison model at coding.

OpenCode and OpenRouter did not describe privacy in exactly the same way

Ox Alpha was also distributed through OpenCode. OpenCode’s preview messaging advertised zero data retention and said the provider would not use customer data for training. 4549

The OpenRouter listing used different wording: it said that prompts and responses were retained by the provider but were not used for training. 4354 Those statements should not be treated as interchangeable. “Not used for training” does not necessarily mean “not retained,” while “zero data retention” is a stronger claim about storage.

For developers, the practical rule is straightforward: treat the access route as part of the privacy decision. A policy advertised for OpenCode should not automatically be assumed to apply when the same model is accessed through OpenRouter or another intermediary. Until the anonymous operator, retention terms, and post-preview conditions are clearer, sensitive source code, credentials, proprietary documents, and secrets should stay out of test prompts. 3248

Why researchers suspected Zhipu AI or Z.ai

Community investigators tried to identify Ox Alpha through behavioral and technical “fingerprints.” Reports pointed to similarities in tokenizer behavior, token counts, video-encoder behavior, response style, and some input-handling patterns. Several analyses connected those clues to Zhipu AI—also known as Z.ai—and its GLM family, with speculation focused on an unreleased multimodal GLM-5.x variant. 1425

That was the leading attribution theory, but it remained circumstantial. Other reporting noted that the apparent serving capacity did not fit the inference neatly, and no official statement from Zhipu AI confirmed that it built or operated Ox Alpha. 14

The distinction matters: a fingerprint can suggest shared components, a related model family, or a common serving configuration, but it is not the same as a vendor announcement. Claims assigning near-certain probabilities to a specific GLM version went beyond what had been independently established in the available reporting.

Xiaomi’s MiMo theory remained an alternative

Some community discussion proposed Xiaomi’s MiMo family as another possible source. The theory was plausible at a high level because MiMo is also associated with multimodal and coding-oriented model development, but the supplied evidence did not establish that connection. 1

At the stated cutoff, the best-supported wording was therefore: Zhipu/Z.ai’s GLM family was the leading hypothesis, Xiaomi MiMo was an alternative theory, and the actual developer was unknown.

Why the launch resembled earlier mystery-model previews

Ox Alpha’s release also recalled earlier anonymous model tests discussed in connection with Chinese AI labs. One frequently cited comparison was Pony Alpha, an earlier anonymous release that reporting associated with Zhipu’s GLM-5 before the model’s public identification. 1454

The resemblance is in the launch pattern: a high-capability endpoint appears under a temporary or anonymous name, developers test it at scale, and the operator can gather real-world feedback before attaching a public brand. That pattern may explain the strategy behind Ox Alpha, but it does not prove that Ox Alpha shared Pony Alpha’s owner, architecture, or training lineage.

Bottom line

Ox Alpha was a real and unusually generous public preview: free access for roughly a week, a 1,048,576-token context window, 131,072-token maximum outputs, and text, image, and video input. Its early 80% DeepSWE result was notable, but it came from only 10 tasks and was not an official leaderboard score.

The strongest origin theory linked the model to an unreleased Zhipu AI/Z.ai GLM-family system, while Xiaomi’s MiMo remained possible but unconfirmed. The privacy picture also depended on the route: OpenCode advertised zero retention, whereas OpenRouter reporting described provider retention without training use. Until the operator identifies itself and publishes durable terms and broader evaluations, Ox Alpha is best treated as a compelling experiment for controlled, non-sensitive testing—not a verified production standard.