An anonymous lab released the Ox Alpha model on OpenRouter, featuring 1M token context, multimodal capabilities, and free access. It was showcased in an agent that autonomously built and verified a Three.js website.
An anonymous lab dropped a model on OpenRouter this week. No name. No paper. No announcement. Just "Ox Alpha". 1M context. Multimodal. Free. Nobody knows who built it. So we did the only reasonable thing: plugged it in as the Brain of Row-Bot and gave it ONE prompt. Research yourself. Build a Three.js website about what you find. Open it in a browser and verify your own work. No hand holding. No retries. I just watched. Phase 1, research: it swept X and the news wires, ingested 15 posts and parsed two primary articles, then cross-checked the specs against its own live runtime config. Verified numbers only: 1,048,576 token context, 131K max output, text + image + video input, native tool calling at ~4.45% error rate, ~50 tokens/sec, 99.99% uptime. It even separated confirmed facts from identity rumors instead of repeating hype. Tokenizer fingerprints point at GLM-5.3, nobody has confirmed anything. Phase 2, build: one single-file HTML page written from scratch. CRT boot terminal, a 14k particle torus-knot hero in raw Three.js, marquee ticker, animated benchmark bars, real community quotes with sources, honest verdict cards listing what DIDN'T hold up too. No frameworks. No templates. Phase 3, self-QA: it launched Chromium, screenshotted section by section and vision-checked its own output like a picky reviewer. Boot overlay clears on schedule. Particles animating. All 8 spec cells render. Capability cards sit in a clean grid. Bars fill correctly. Zero rendering errors across every pass. ~20 tool calls spanning research, codegen, browser automation and visual QA. One session. One prompt. Honest part: folks are reporting it's slow under load (~11.6s median agent-turn latency tracks) and prompts get retained by an anonymous provider, so never send secrets to a stealth preview. Still. Step back and look at what happened. An unidentified frontier model researched itself, designed its own showcase and QA'd it end to end inside an open source agent harness. Benchmarks are curated highlights. This was the whole job, done live, with receipts. And here's the kicker: you don't need to wire up APIs yourself. Ox Alpha ships in Row-Bot right now as a first class model pick (both the OpenRouter stealth route and the free OpenCode Zen unlimited tier). Pick it in Settings > Models and run your own gauntlet before the free window closes.
The anonymous AI model Ox Alpha processed 26 trillion tokens on OpenCode in four days, breaking OpenRouter's launch record, and was introduced as a free, multimodal model with a large context window through an open-source coding agent.
Ox Alpha is described as the most popular free reasoning model on OpenRouter, featuring a 1M token context window for coding and agentic tasks, and can be tested via Pi or Hermes Agent.
A mysterious AI model named Ox Alpha has been released on OpenRouter, leading to widespread speculation about its developers, with theories pointing to Chinese AI companies like Z.ai or other entities.
Ox Alpha has been released for free on OpenRouter and is being praised for its impressive performance, with users comparing it favorably to other top models like GPT-5.6 and Fable 5.
A free and open AI model named Ox Alpha, with about 1 million Token context and multimodal capabilities, suddenly jumps to second place on OpenRouter's weekly chart, earning high praise from Stripe CEO Patrick Collison, and potentially disrupting the AI industry's valuation logic.