On August 20, 2026, a free model called "Ox Alpha," developed and operated by an anonymous third party, appeared on OpenRouter and OpenCode. Its 1,048,576-token context window and coding performance drew attention first. However, a premise that development teams should verify before anything else is that, even when sending code to the same Ox Alpha, who handles the data and under what conditions it can be retained or used differs depending on the route used to access it.
Code agents read source code and also touch repository configurations, logs, and test results. Sometimes they wait for instructions near authentication credentials. OpenRouter describes itself not as the developer, owner, or provider of Ox Alpha, but as an intermediary that routes requests, and it does not disclose the operator's name or location. The unit that users need to verify is not "the model named Ox Alpha" but rather the access route—whether via OpenRouter or via OpenCode.
A 1,048,576-Token Spec, and a Benchmark Recast at 58.4%
OpenRouter's model ID is stealth/ox-alpha, and the model page lists its release date as August 20. Both input and output pricing are $0, with a context window of 1,048,576 tokens and a maximum output of 131,072 tokens. It accepts text, image, and video input. Reasoning is always enabled, with strength selectable among low, high, and max, defaulting to max.
This large input window is broadly consistent with measurements from a published third-party investigation. LuD1161's investigation successfully retrieved information embedded at the beginning, middle, and end of the input in all three trials at a state of 934,221 tokens. It reported the acceptable input boundary at approximately 1.00M to 1.005M tokens and the output cap at 131,072 tokens—not far from the stated specifications.
However, the "8 out of 10" impression that spread early on cannot simply be taken as representing the model's overall capability. In an independent run of DeepSWE published by MatchaOnMuffins, the model solved 66 out of 113 problems, a result of 58.4%. The run took 20 hours and 39 minutes, with each problem attempted once. Since this was not an official leaderboard run, it cannot be placed on the same footing as averages from multiple runs or results from different execution environments.
The execution log also records that in 11 problems, tool calls failed to return three times in a row, causing the run to stop with a RepeatedFormatError. This corresponds to 9.7% of all 113 problems. The ability to solve coding problems and the ability to reliably complete long agent runs need to be measured separately.
A GLM-5-Generation Fingerprint, But Identity Unconfirmed
As of August 24, neither the developer nor the operator of Ox Alpha has been officially confirmed. Nevertheless, published investigations are attempting to narrow down candidates based on the model's behavior. LuD1161 combined two lines of investigation, cross-referencing over 600 requests and approximately 13.5M input tokens.
The most notable finding is a tokenizer comparison. 44 identifier strings matched the GLM-5-generation tokenizer at a rate of 44/44, compared to 42/44 for GLM-4.x. The investigation also cross-checked upstream errors returned in Chinese, reasoning-strength controls, and context-boundary behavior, assessing attribution to the GLM-5 family as high-confidence and attribution to a specific checkpoint as moderate-to-high confidence.
However, this is a black-box investigation observing input and output from the outside. A 44/44 match is a strong clue, but it does not prove that the model weights are identical or trace the origin to a specific checkpoint. An assessment that the model resembles GLM-5 does not identify Ox Alpha's operator or where processing takes place.
As long as the anonymous provider's name cannot be confirmed, a performance comparison table alone cannot settle a procurement decision. It is necessary to verify the operating entity and processing location, and to review data retention and reuse as a separate contractual condition.
Data Descriptions Diverge Between OpenRouter and OpenCode
OpenRouter's model page states that the provider retains input and output but does not use them for training. The retention period is not specified. Reading this description alone, one could take it to mean there is no training use at all.
OpenCode Zen offers the same Ox Alpha as x-preview-f-free, providing it for $0 via a chat API. Its individual description and privacy section state that the provider follows a zero-retention policy and does not use data for training. All models are said to be hosted within the United States, and Ox Alpha is not listed among the exceptions.
The gateway, contract, and hosting policy differ between the OpenRouter route and the OpenCode route. Even if the model weights are identical, the terms applied to the input data users send would differ.
| Access route | Disclosed data handling | Location/retention period disclosure |
|---|---|---|
| OpenRouter | Model page states the provider retains input and output but does not use them for training | Retention period not specified. Operator name and location also undisclosed |
| OpenCode Zen | Ox Alpha Free is zero-retention and not used for training | States all models are hosted within the U.S. Ox Alpha is not on the exception list |
The terms of use for a model with the same name cannot be lumped together simply by model name. OpenCode's zero-retention and U.S.-hosting claims are policies the company itself has disclosed, and there is no information available for third parties to audit regarding the name of the anonymous upstream provider. This limitation remains.
Within OpenRouter Itself, "No Training Use" Conflicts With the Stealth Terms
Within OpenRouter, there is even a discrepancy between descriptions along the same route. While the model page states that data is "retained but not used for training," the Stealth Program EULA, updated July 6, 2026, states that in exchange for providing anonymous models for free, user content is collected, shared, and used for training, evaluation, and improvement.
The EULA grants OpenRouter a worldwide, irrevocable, perpetual license to store, use, and distribute the content, among other things. It also permits sublicensing to the anonymous provider for training, evaluation, and improvement purposes. When passing data to the provider, user identifiers are hashed, and the provider is prohibited from re-identifying users. This is not a guarantee that personal or confidential information is stripped from the input body, nor is it a measure that narrows the purposes for which input and output can be used.
Even reading the model page and the EULA together, one cannot definitively determine whether Ox Alpha's input and output are used for training. The actual retention period and processing location also remain unknown. As long as the discrepancy between the brief model description and the legal contract document remains unresolved, users cannot pin down how their submissions will be handled after sending them.
Free provision can serve as a means of quickly gathering real-world traffic and failure patterns. However, what purpose Ox Alpha's provider is pursuing has not been disclosed. Users should not fill in that purpose with guesswork, and instead should determine the scope of data they may send based on the broadest possible usage license.
Before Adopting a High-Performing Anonymous Model: Who Are You Sending What To?
If a company is going to evaluate Ox Alpha, the starting point is to test its capability and agent stability separately, under conditions where confidential code, configurations, and logs are not sent. The 58.4% DeepSWE result comes from an independent run with public records, but it is no substitute for the early "8/10" figure or an official leaderboard. It would be premature to take a performance evaluation obtained this way and directly link it to a decision to send production data.
Next, verify the anonymous provider's identity, processing location, and retention period. Terms of use for training and evaluation, along with contact information in case of an incident, are also necessary. These gaps should not be filled in with model performance. Until OpenRouter's terms hierarchy and the actual processing conditions are clarified, it becomes necessary to limit the information sent through that route to a scope where such uncertainty is acceptable.
Even if a company chooses OpenCode, the zero-retention and U.S.-hosting claims are disclosures made by that company itself, and do not mean that the anonymous upstream provider has been independently verified as well. The conditions under which a company can send production code are determined not by the 1,048,576-token context length, but by who can vouch for the provider's name and the data-processing terms.