A stealth model called Ox Alpha went live on OpenRouter on August 20, and its provider has declined to say who built it. The listing gives away the specs but not the source: a 1,048,576-token context window, a 131,072-token maximum output, and nobody pays to use it: input tokens and generated replies both cost nothing on OpenRouter’s listing. That combination, frontier-scale context at no cost, is the actual news here, not the anonymity gimmick.
OpenRouter is explicit about its own role: it routes requests to Ox Alpha and is “not its developer, owner, or provider.” The model is built and run by a third party that has chosen to stay unnamed during this preview period. Nothing on the page hints at who that party is, and nothing here should be read as a guess.
Ox Alpha is pitched for coding, long-horizon agentic work, and production workloads rather than casual chat. Feed it a prompt built from words, pictures, or a video clip, and whatever comes back is written language only. It supports tool calling (tools and tool_choice) plus response_format for JSON output, though without JSON-schema enforcement. That last detail matters for anyone building structured-output pipelines: the model will try to return valid JSON, but it will not guarantee a schema match the way stricter API modes do.
Performance numbers on the page are concrete rather than promotional. Median throughput is 31 tokens per second and median latency is 3.31 seconds, both listed as the best figures across providers serving the model, since Ox Alpha currently has only one provider. Uptime over the prior three days measured 99.99 percent, with availability at 98.93 percent for the model and 98.99 percent for OpenRouter’s routing layer over the last 24 hours. None of these numbers come with a comparison benchmark against named competitors: the release announcement does not include independent evaluation results.
The data-retention term deserves its own read. OpenRouter’s listing states plainly that the anonymous operator keeps a copy of everything you send it, and everything it sends back, but has committed to keeping that history out of any future training run. Retained but not trained on is a narrower promise than deleted, and it is the one operators should weigh before routing anything sensitive to a free endpoint whose owner they cannot identify or contract with directly. Everything else about usage falls under OpenRouter’s Stealth Model Terms, a separate document from the standard terms that govern named providers.
Traffic data on the page shows which tools are already sending volume to Ox Alpha. Anthropic’s Claude Code leads at 68.2 billion tokens, followed by Hermes Agent, an open-source persistent agent from Nous Research, at 64.8 billion. Two newer entrants, Oh-My-Pi and a DeepSeek Harness multimodal bridge, follow at 56.8 billion and 49.9 billion tokens, with ZCode rounding out the top five at 33.7 billion. That a coding-agent workload dominates the traffic mix lines up with how the model is positioned: long-context, tool-using, built for sustained sessions rather than single-turn answers.
Free access to a million-token, tool-calling, multimodal-input model is unusual enough on its own that the provider’s silence about their own identity is a secondary story. Teams evaluating Ox Alpha for production agentic pipelines should treat the retention terms as a hard constraint, not a footnote, and should confirm the lack of JSON-schema enforcement will not break existing structured-output tooling before they route real traffic to it.
Per OpenRouter’s Ox Alpha model page, published August 20, 2026.
