A free, anonymous AI model just landed on OpenRouter. The launch tactic is the story.

A free, anonymous AI coding model called Ox Alpha surfaced on OpenRouter this week, and the story that actually pays off is not who built it but the tactic that put it there. The model appeared Thursday, listed by an unnamed third-party provider, and was quickly described by OpenCode, an open-source AI coding agent, as free for a week with "near unlimited usage." OpenCode said the provider had capacity for 100 trillion tokens per day, roughly 100 times the AI tokens Visa said it uses in a month, a figure no one outside the provider has verified.

OpenRouter, a major routing platform that lets developers swap between AI models, describes Ox Alpha as "a reasoning model designed for coding, sustained agentic work, and production workloads," suited for "long-horizon software engineering, complex reasoning, and workflows that combine text with visual context." Agentic work means multi-step tasks where the model acts on its own, like running tests or editing files. KIE.AI, a model directory, lists Ox Alpha as supporting a 1M-token context window and video input, a wide enough frame to hold a long code session or a short clip.

Stripe co-founder Patrick Collison tried the model and posted on X that it was "very impressive," a remark from one CEO, not a consensus. Early speculation, originally surfaced by Wccftech, pointed at Z.ai, the Chinese lab behind the unreleased GLM-5, on the strength of tokenizer behavior (the way a model breaks text into processable units) and response patterns. Wccftech also raised a competing hypothesis: that the model could be from Microsoft's MAI family. By Saturday, Andrew Curran wrote on X that the Z.ai theory had been the leading one on Friday night, but by Saturday morning "people seem less sure of anything."

That hedging is the story. Anonymous model drops have stopped being anomalies. Last year, Z.ai tested GLM-5 under the name "Pony Alpha" before going public, a pattern that is now standard inside the Chinese AI ecosystem. Moonshot's Kimi K3, a 2.8-trillion-parameter open-weight model whose trained parameters are released publicly, dropped in July and quickly drew attention in Silicon Valley. DeepSeek and Zhipu, two of the better-known Chinese labs, have spent the past year releasing capable open-weight models at price points well below their US counterparts. A frontier-grade, free, anonymous release now reads as a probe, not a glitch.

A lab wants developer attention before a formal launch. It routes a model through a platform like OpenRouter under a placeholder name, lets open-source coding agents wire it up, and watches the developer class form a verdict in real time. If the verdict is good, the lab has social proof, a workload footprint, and a free week of telemetry before any marketing spend. If the verdict is bad, the model vanishes and the lab learns in public without its name attached. Either outcome is useful, and the cost is a few days of free inference.

The part that is harder to verify is the 100-trillion-tokens-per-day claim. That number comes from OpenCode's X post, not from a primary provider disclosure, and the only public handle on it is OpenRouter's own announcement tying Ox Alpha to its stealth category. No independent benchmark, no model card, no paper.

Identity will likely get settled in days or weeks. Whoever is behind Ox Alpha will either reveal itself or be unmasked by a tokenizer tell or a server log. The pattern, by contrast, is now structural. Business Insider framed the moment as a mystery, and it is one, but anonymous drops are now the launch tactic in the China-US model race, and Ox Alpha is the cleanest example so far.