Anonymous 'Ox Alpha' stealth model lands on OpenRouter with 1M context, free preview

OpenRouter has begun routing traffic to Ox Alpha, a reasoning model pitched at coding, long-horizon agentic tasks, and production workloads. The provider behind it has chosen to stay anonymous for the preview — OpenRouter stresses it merely forwards requests and is neither the developer nor operator. The model is multimodal on input, accepting text, images, and video while returning text only, and it supports function calling and JSON-mode output (though without schema enforcement). Its headline specs are a roughly 1.05M-token context window with up to 131K completion tokens, and it’s currently free to use.

The usage data is the more interesting tell. Early traffic is dominated by agentic coding tools, led by Anthropic’s Claude Code at 22B tokens, followed by Nous Research’s Hermes Agent at 17.4B and ZCode at 10.6B — a signal that the target market is sustained, tool-heavy software engineering rather than one-off chat. Reported performance is modest at around 34 tokens/second throughput and 3.74s latency, with uptime near 99.99% over three days.

On data handling, the terms are a mixed bag worth flagging: prompts and completions are not used for training, but they are retained by the anonymous provider, and all other use falls under OpenRouter’s Stealth Model Terms. Stealth launches like this are typically a lab dogfooding an unreleased model under a codename before a formal reveal, so the combination of a free tier, a million-token window, and an undisclosed operator is best treated as a preview with unknown provenance. Released August 20, 2026.