A model with no name attached to it just posted numbers that would make most AI labs blush, and nobody in Silicon Valley can say for certain who built it. Since last Thursday, developers have been testing a system called Ox Alpha — sometimes typed as “0x Alpha” in social posts — after it quietly appeared on the model marketplace OpenRouter and inside the coding tool OpenCode. This kind of anonymous AI model release has become an increasingly common industry move, but Ox Alpha’s scale and secrecy have turned it into one of the more talked-about mysteries in AI this month.
Summary
Key takeaways
- Ox Alpha is a stealth AI model released on OpenRouter and OpenCode, listed only as coming from “a third-party provider who has chosen to remain anonymous during this preview.”
- It claims a context window of 1,048,576 tokens (roughly one million), a maximum output of 131,072 tokens, and an alleged daily capacity of 100 trillion tokens.
- The model supports text, image, and video input and is pitched for coding, long-running agent tasks, and production workloads.
- It is free during a limited preview on OpenCode Zen and included with the paid OpenCode Go subscription.
- Community figures suggest it processed roughly 7.1 trillion tokens in its first week across about 134,000 individual users, though none of the capacity or performance claims have been independently verified.
Ox Alpha’s Anonymous Debut and Ambitious Specs
Ox Alpha showed up with no company name, no press release, and no confirmed origin — just a listing describing it as coming from an anonymous third-party provider. OpenRouter’s own description calls it a “reasoning model designed for coding, sustained agentic work, and production workloads,” suited to “long-horizon software engineering, complex reasoning, and workflows that combine text with visual context,” according to Business Insider. That framing tells developers exactly what the model is built to do, even if it says nothing about who built it.
Multimodal Inputs and Programming Focus
Beyond the mystery, the Ox Alpha AI capabilities on paper are broad. The model accepts text, image, and video input, and it’s aimed squarely at programming tasks and agent-style workflows that run for extended stretches rather than single quick answers. That positioning matters because it targets the exact audience — developers running long, expensive agent sessions — most likely to stress-test a model hard and talk about it publicly.
Massive Capacity Claims Remain Unverified
The specifications claimed for Ox Alpha are striking: a context window of 1,048,576 tokens, roughly one million, a maximum output of 131,072 tokens, and an alleged capacity of 100 trillion tokens per day. For scale, Business Insider noted that figure is roughly 100 times the number of AI tokens Visa has said it burns through in an entire month. None of these numbers have been independently confirmed, and capacity is not the same thing as quality — a model can swallow an enormous prompt and still lose track of what’s inside it. Independent benchmarks from groups like Artificial Analysis or rankings on Arena.ai are still pending.
How Developers Are Accessing the Model
Right now, testing Ox Alpha costs nothing, which is the single biggest reason it spread so fast. The model is available through two routes on the OpenCode AI platform: OpenCode Zen, where tokens are free during the preview though billing details are required, and OpenCode Go, a monthly subscription that includes the model at no extra charge. OpenCode itself is an open-source coding agent built by the team behind the SST infrastructure framework, and unlike rivals such as Claude Code or Cursor, it isn’t tied to one in-house model — developers can swap between providers freely, which is exactly what made it an appealing launchpad for a stealth release.
The response has been fast and loud. According to community figures, Ox Alpha in its initial seven days, approximately 134,000 distinct users collectively consumed around 7.1 trillion tokens. Stripe CEO Patrick Collison reportedly tried the model and called it “very impressive” in a post on X, an early signal that helped fuel the buzz even before anyone knew who was behind it.
The Hunt for Ox Alpha’s Creator
With the lab staying silent, developers have turned to digital forensics — picking apart tokenizer behavior, error messages, video token counts, and phrasing patterns for clues. So far, that detective work hasn’t produced a confirmed answer.
Four Competing Theories
Zhipu AI, commonly referred to as Z.ai, is considered by many to be the leading contender. Token fingerprinting analysis has seemingly revealed toward the company’s GLM family across multiple test runs, and Wccftech noted that Z.ai previously tested its GLM-5 model anonymously under the codename “Pony Alpha,” giving the theory some precedent. Zhipu also runs GPU clusters large enough to plausibly support the giveaway, though the sheer volume of free compute cuts against the idea.
Microsoft entered the conversation because Ox Alpha reportedly uses the cl100k_base tokenizer, an encoding developed by OpenAI that also shows up in Microsoft’s Phi and MAI model lines — pointing, on this theory, toward an unreleased version of MAI 2. Google became a suspect after recently shipping Gemini 3.7 Flash while a long-rumored Gemini 3.5 Pro remains missing, though a vision test that Gemini models usually pass and Ox Alpha reportedly fails works against that theory. A fourth trail leads to Cursor and SpaceX, with speculation centering on Cursor’s next coding model, Composer 3, said to have trained on SpaceX-linked supercomputer infrastructure — an argument built mainly around the fact that few players could plausibly give away hundreds of trillions of tokens a day.
Even the online sleuthing hasn’t settled anything. AI analyst Andrew Curran wrote on X that GLM was the leading theory Friday night, but by Saturday morning, “people seem less sure of anything.” The uncertainty itself has become part of the story, feeding a wave of speculative posts, videos, and articles that keep the model in front of new testers every day.
Why Labs Are Embracing Anonymous AI Model Releases
Ox Alpha isn’t an isolated stunt — it’s the latest entry in a pattern that’s become a genuine playbook. Google’s Nano Banana image generator followed the same path: released anonymously for testing under a throwaway codename, it climbed community rankings before being revealed as a Google product, eventually spawning Nano Banana Pro and Nano Banana 2. OpenRouter alone has hosted roughly a dozen similarly cloaked listings. Quasar Alpha and Optimus Alpha proved to be early-stage iterations of GPT-4.1; Horizon Alpha and Horizon Beta were early GPT-5 checkpoints; Hunter Alpha and Healer Alpha were later linked to Xiaomi, and Owl Alpha to Meituan. Other codenames, including Sonoma Dusk, Polaris Alpha, and Aurora Alpha, were never definitively attributed at all.
This matters beyond the guessing game. Chinese labs like Zhipu, DeepSeek, and Moonshot AI have been closing the gap with U.S. rivals at a fraction of the cost, often through open-weight releases — Moonshot’s Kimi K3, a 2.8 trillion-parameter open-weight model released in July, is a recent example that drew serious attention in Silicon Valley. An anonymous AI model release lets any lab, Chinese or American, gather real-world usage data and feedback without attaching its brand to a product that might underdeliver.
The Real Strategy Behind the Free Tokens
The most revealing part of this launch may not be the technology at all — it’s the pricing. Giving away tokens on both input and output, for a class of model where long agent sessions can otherwise run up three-figure bills fast, is a powerful incentive to try something with no proven track record. Combine that with a massive context window and a ticking clock on the free preview, and you get the kind of urgency a conventional product announcement rarely manages to create anymore.
That urgency is doing double duty for whoever built it. Every tester who posts results pulls in the next wave of testers, generating real-world usage data and feedback the lab would otherwise have to pay dearly to collect. And because nobody knows the source, every theory floated online — Zhipu, Microsoft, Google, Cursor — generates another round of posts, videos, and articles, extending the model’s reach for free. If Ox Alpha does turn out to be the next Nano Banana, this brief window of anonymity may end up remembered as its cheapest and most effective marketing campaign.
FAQ
What is Ox Alpha?
Ox Alpha is an anonymous AI model released on the OpenRouter and OpenCode platforms, positioned mainly for programming and long-running agent tasks.
What are Ox Alpha’s key capabilities?
Ox Alpha supports multimodal inputs including text, images, and video, claims a context window of about one million tokens (1,048,576), and an alleged daily token capacity of 100 trillion.
Who developed Ox Alpha?
The lab behind Ox Alpha remains unknown. Speculation points to Zhipu AI, Google, Microsoft, or Cursor and SpaceX, but none of these theories has been confirmed.
How can developers access Ox Alpha?
Ox Alpha is available free during a limited preview through OpenCode Zen and is also included with the paid OpenCode Go subscription.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
