Ox Alpha Lands on OpenRouter as a Free 1M-Context Stealth Model for Coding and Agentic Work

OpenRouter has listed Ox Alpha, an anonymous “stealth” reasoning model aimed squarely at coding, sustained agentic work and production workloads. The unusual combination is hard to miss: a 1,048,576-token context window, multimodal input support for text, images and video, function calling and JSON output support — and, at least during this preview, zero token pricing.

Released on August 20, 2026, Ox Alpha is presented as a model for long-horizon software engineering and complex reasoning workflows that need to carry substantial context forward. OpenRouter says the model can accept text, image and video inputs and returns text, while supporting tools and tool_choice for function calling. It also supports response_format for JSON output, although without JSON-schema enforcement.

The bigger story is the model’s anonymity. OpenRouter explicitly says Ox Alpha is developed and operated by a third-party provider that has chosen not to identify itself during the preview. OpenRouter is acting as the routing platform rather than the developer or owner. Prompts and completions are retained by the provider, though OpenRouter says they are not used for training; users should therefore treat the privacy and data-retention terms as a meaningful part of the trade-off.

On raw headline specifications, the preview is unusually aggressive. The model exposes a 1M-token context window and supports up to 131,072 completion tokens. OpenRouter currently lists the model as free for both input and output tokens. Its recent platform metrics show roughly 24 tokens per second throughput and a P50 latency of 5.47 seconds, with three-day availability around 99.50% and uptime around 99.99%.

Those numbers make Ox Alpha interesting for agentic coding systems that would normally have to aggressively summarize or discard context. A million-token window can make it practical to keep large codebases, logs, documentation and long-running task histories in a single working context. Tool calling and multimodal inputs also position it for workflows that combine repository work with screenshots, diagrams, UI inspection or video-derived context.

But this is still a preview, and the anonymity changes how the model should be evaluated. There is no public developer identity to use as a proxy for model lineage, training approach or long-term support. Free pricing may also be temporary, and current performance statistics are platform observations rather than guarantees. For production systems, the sensible approach is to benchmark Ox Alpha against known alternatives on the actual workloads that matter: coding accuracy, long-context retrieval, agent reliability, tool-use behavior and latency under load.

The signal worth watching is not simply that another large-context model has appeared. Ox Alpha packages several features that are normally expensive — long context, reasoning, multimodal inputs and agent tooling — into a free preview with an anonymous provider behind it. If the quality holds up under independent evaluation, it could become a useful testbed for long-running coding agents. If it does not, the episode still illustrates how quickly the competitive baseline for agent-oriented models is moving.

For now, Ox Alpha is best understood as a high-capability but still opaque preview: technically ambitious, operationally interesting, and deserving of careful benchmarking before anyone treats the headline specifications as proof of production readiness.