Ox Alpha: A Free 1M-Token Coding Model with Unresolved Data Risks

A frontier coding model with a 1,048,576-token context window appeared on OpenRouter on August 20, 2026, with no company name and conflicting data-handling policies. For AI builders evaluating long-context code reasoning, the model offers a rare stress-test opportunity, but the unresolved provider identity and contradictory retention terms make it risky for sensitive work.

The model, listed as stealth/ox-alpha, is free through OpenRouter and via the OpenCode agent platform for about one week. OpenRouter describes it as a reasoning model for coding and sustained agentic work. The context window alone lets you load an entire large codebase into a single prompt, which is a different category of tool than typical 8K-200K token models. Throughput sits around 29 tokens per second at this scale.

Forensic Signals Point to Zhipu AI

Developer forensics published by August 22 provide the strongest available evidence about the provider. Tokenizer fingerprinting on 30 probe strings covering 14 writing systems matched GLM-5.3's tokenizer exactly, with a constant 75-token offset from a routing wrapper. Video encoder analysis matched Zhipu AI's GLM-5V-Turbo across three independent design choices. Most concretely, a Java stack trace from a deliberately malformed request exposed internal class names mapping to Z.ai's API route at open.bigmodel.cn. Error-code dialect matched Z.ai-hosted GLM models, not the same model served by a different operator. No lab has officially confirmed, but the evidence is compelling.

The Real Risk: Conflicting Data Policies

The per-model listing on OpenRouter states that prompts and completions are retained but not used for training. However, the OpenRouter Stealth Model EULA grants OpenRouter and the unnamed provider rights to use user content for training, evaluation, and improvement. These two documents conflict, and the published materials do not explain how the conflict is resolved in the developer's favor.

The OpenCode route for Ox Alpha carries a different policy, claiming zero days of retention and no training use. But that policy applies to the same anonymous provider whose identity and jurisdiction remain officially unknown. Two routes to the same model, two different stated data policies.

Compliance and Regulatory Exposure

If the forensic evidence holds and Zhipu AI's Z.ai infrastructure is confirmed, a distinct category of legal risk applies. Z.ai operates under China National Intelligence Law Article 7, which requires all Chinese organizations to cooperate with state intelligence work. Z.ai is also on the US BIS Entity List for advancing China's military modernization through AI. US enterprises routing proprietary code through this infrastructure should assess compliance implications.

EU Regulation 2026/1744 deferred the high-risk AI compliance framework that would have required formal data processing agreements with named AI providers. However, Article 50 transparency requirements for AI-generated content remain in effect. European enterprises should consult current guidance rather than assuming regulatory clarity.

What Builders Should Do

For evaluation on non-sensitive inputs, Ox Alpha is an unusual research opportunity. The free window closes around August 27, and based on prior stealth model patterns (Xiaomi MiMo-V2-Pro, Meituan LongCat-2.0), a formal reveal with named terms is likely soon. For any prompt containing proprietary code, withhold your data until a named, auditable provider with clearly binding data-processing terms is available. Treat the EULA conflict as unresolved and retention as nonzero until proven otherwise.