Anonymous Model Ox Alpha Offers a Free 1M Context Window, but the Provider Retains Prompts and Outputs

OpenRouter’s anonymous preview model Ox Alpha has recently become a popular testing endpoint among agent developers. It is currently priced at $0 per million input tokens and $0 per million output tokens, with a context window of approximately 1.05M tokens, and is positioned for coding, long-horizon agent tasks, and workflows incorporating visual content. OpenRouter stresses that it is only the router: the model is developed and operated by a third party that has chosen to remain anonymous, with no public disclosure of its architecture, parameter count, training sources, or official product name. [OpenRouter’s provider page](https://openrouter.ai/provider/stealth) is currently the most direct official record.

The technical interface is fairly comprehensive. Public information lists the model ID as `stealth/ox-alpha`. It can be accessed through OpenAI-compatible, Responses, and Anthropic Messages interfaces; accepts text, image, and video inputs; supports a maximum output of 131,072 tokens; and offers tool calling, structured outputs, and continued reasoning state. These specifications are sufficient to ingest a large codebase, documentation, issues, and test logs in a single request, which also explains why it was quickly integrated into workflows such as OpenCode and Hermes Agent. However, the widely circulated ten-task DeepSWE test used an extremely small sample and cannot establish whether Ox Alpha outperforms known frontier models. [AI Primer’s technical overview](https://www.ai-primer.com/engineer/stories/ox-alpha-openrouter-release) also characterizes the result as a smoke test.

The more consequential change concerns the data policy. Early promotional claims described the testing as having zero data retention, but the current model page explicitly states that the anonymous provider retains prompts and completions, promising only that they will not be used to train models; no retention period has been disclosed. OpenRouter’s default policy of not storing content does not mean that downstream model providers also follow zero data retention (ZDR). Its documentation allows customers to set `provider.zdr: true` to force routing exclusively to zero-retention endpoints; the request should fail if no eligible endpoint is available. [OpenRouter’s ZDR documentation](https://openrouter.ai/docs/guides/features/zdr) explains the distinction between these two policy layers.

Ox Alpha is therefore suitable for compatibility, long-context, and tool-calling experiments using public code and synthetic data, but users should not submit private repositories, credentials, customer records, or regulated data. Engineers should watch for disclosure of the provider’s identity, alignment of model-specific terms, and pricing and capacity after the free period ends. Until then, the anonymous origin and data-retention policy are deployment blockers in their own right, not merely details buried in a legal page.