Ox Alpha: What We Know About the Anonymous AI Model and the Zhipu Theory

OpenCode also offered access through its free tier. Its announcement described the preview as having generous limits and “near unlimited usage,” with a claimed capacity of 100 trillion tokens per day. That figure is an operator claim, not an independently audited throughput measurement.

The headline result: 80% on a small DeepSWE subset

Developer Ben Davis tested Ox Alpha on 10 tasks drawn from the DeepSWE software-engineering benchmark. It reportedly passed eight of them, producing an approximately 80% result. On the same limited comparison, Claude Fable 5 scored 65% and GPT-5.6 Sol scored 52%.

That is an attention-grabbing result, but it should not be described as an official DeepSWE victory. DeepSWE v1.1 contains 113 original, long-horizon software-engineering tasks across 91 repositories, so the 10-task experiment covered only a small fraction of the benchmark.

The comparison also had important methodological limitations:

• The sample was small and therefore vulnerable to task-selection effects and high variance.
• The compared models did not necessarily use the same number of rollouts or attempts.
• Ox Alpha had not been evaluated under the official leaderboard protocol or added to the public leaderboard.
• A score from one researcher’s subset is not directly equivalent to a full-benchmark pass rate with standardized evaluation conditions.

The public leaderboard snapshot available at the time placed Claude Opus 5 first at 73.6%, GPT-5.6 Sol at 72.7%, and Claude Fable 5 at 69.7%. The careful conclusion is therefore narrower: Ox Alpha looked very strong in Davis’s 10-task experiment, but the result does not establish that it was the best general-purpose coding model.

Why analysts suspect Zhipu AI

The leading attribution theory connects Ox Alpha to Zhipu AI, also known as Z.ai, and its GLM model family. The case is based on behavioral fingerprinting rather than a company announcement.

One reported test compared Ox Alpha’s tokenization behavior with GLM-5.3 across 25 prompts. The underlying token counts reportedly matched after accounting for a constant 75-token wrapper. Separate video tests reportedly found matching visual-token behavior, including similarities in frame sampling, duration scaling, and resolution handling.

Other reported similarities include response style, API behavior, and how the model handles audio input. Taken individually, any one of these clues could be explained by a shared wrapper, common infrastructure, or imitation. Together, analysts argue, they form a stronger cluster consistent with Zhipu’s unified multimodal GLM line.

Zhipu’s previous use of anonymous or stealth-style releases, including the Pony Alpha moniker, adds circumstantial context. It does not prove that Zhipu created Ox Alpha.

What remains unverified

No company had publicly claimed Ox Alpha during the period covered by the reports, and Zhipu had not confirmed the attribution. Speculation has included specific GLM variants, but the available evidence does not establish whether Ox Alpha is an unreleased model, a production variant, a fine-tuned system, or an independently operated model using related infrastructure.

The defensible description is therefore: Ox Alpha is likely GLM-derived and may be connected to Zhipu, but its creator and exact model identity were unverified. Confidence percentages circulated by individual analysts should not be treated as official identification.

The privacy and data-retention split

Ox Alpha’s privacy language is one of the most important practical details for developers. The model’s OpenRouter listing said that prompts and completions are retained by the underlying provider but are not used for training.

That statement is different from OpenRouter’s general data-collection policy, which says OpenRouter itself does not store prompts or responses unless users opt into input/output logging. OpenRouter also documents provider-level policies separately, meaning the routing layer and the model provider can have different retention rules.

OpenCode, meanwhile, advertised Ox Alpha with “zero data retention.” That promise may describe OpenCode’s own handling of requests, but it does not automatically cancel the provider-retention disclosure attached to the OpenRouter route. In other words, “OpenCode does not retain data” and “the underlying model provider retains prompts and completions” can both describe different stages of the same request path.

Until the providers publish a single, contractually clear data-processing policy, the prudent assumption is that the model provider may retain submitted content. Developers should avoid sending proprietary source code, credentials, personal data, or regulated information merely because the service is free or is described as not training on prompts.

Bottom line

Ox Alpha was notable for three reasons: a free short-term preview, unusually large advertised context and multimodal capabilities, and a striking early coding result. But the evidence supports a measured interpretation rather than the claim that an unknown lab definitively defeated the established frontier.

The 80% DeepSWE figure came from eight successful tasks out of 10, not the full 113-task benchmark. The public leaderboard still showed Claude Opus 5 ahead, and Ox Alpha had not completed the official evaluation process.

Its technical fingerprints make a Zhipu/GLM connection the leading explanation, yet no public confirmation established authorship. For experimentation, Ox Alpha was an intriguing model to test. For production systems or sensitive code, its anonymous ownership, provider-level retention, and unresolved identity were reasons to proceed carefully.