OpenRouter tests free Ox Alpha with a 1M-token context window

• Ox Alpha is a free OpenRouter stealth preview with a 1M-token context window and text, image, and video input, as OpenRouter's launch post describes.
• The data-policy claims differ: opencode's launch post called the test zero data retention, while OpenRouter's privacy note said the provider does not train on prompts or completions; the current model page says the provider retains both.
• The anonymous provider has exposed stealth/ox-alpha through OpenAI-compatible and Anthropic-compatible APIs, with tools, structured output, and reasoning controls documented on OpenRouter's listing.
• A small DeepSWE sample from davis7 put Ox Alpha at 80% across 10 tasks, ahead of the poster's 65% Fable and 52% GPT-5.6-sol runs.

The model listing gives the preview a 131,072-token output cap. Hermes Agent added it immediately, while a LocalLLaMA post turned the missing provider name into the first community investigation.

OpenRouter exposes Ox Alpha as stealth/ox-alpha. Its quick-start documentation uses the standard OpenAI-compatible endpoint, while the model listing also documents Responses and Anthropic Messages interfaces.

The documented contract is unusually expansive for a free stealth test:

• Context: 1,048,576 tokens, with a 131,072-token maximum output.
• Modalities: text, images, and video in; text out.
• Controls: reasoning, max_tokens, temperature, top_p, top_k, tools, tool_choice, and response_format.
• Reasoning continuity: OpenRouter says clients can pass the complete reasoning_details array back on later turns.

The privacy language is the material caveat. opencode's launch post advertised “Zero Data Retention,” and OpenRouter's follow-up says the provider does not train on prompts or completions.

The current Ox Alpha page says prompts and completions are retained by the anonymous third-party provider and are not used for training. The listing supplies no retention period, and says OpenRouter routes requests rather than operating the model.

OpenRouter identifies the developer only as a third-party provider that chose to remain anonymous during the preview. The provider question also anchored a LocalLLaMA post shortly after the model appeared.

AiBattle_ first floated a Xiaomi MiMo connection from a token behavior comparison and Xiaomi's earlier anonymous OpenRouter tests. A later follow-up instead relayed unverified reports of GLM-like tokenizer behavior and error messages. Neither OpenRouter nor a model developer has attached a name to Ox Alpha.

davis7 ran Ox Alpha through 10 DeepSWE tasks and reported 80%, with the explicit warning that such a small subset could have substantial variance.

The table records eight passes and two misses, versus the same poster's 65% for Fable, 62% for GLM-5.3 and Grok 4.6, and 52% for GPT-5.6-sol. It is an early smoke test, not a published benchmark result.

opencode described the offer as free for the next week, with generous rate limits, near-unlimited usage, and 100T tokens per day of claimed capacity.

OpenRouter's listing currently prices input and output at $0. Teknium said Ox Alpha was added to Hermes Agent through opencode and OpenRouter, giving the preview an immediate agent surface as well as direct API access.