Sixty-three percent. That’s where a Manifold prediction market landed on Z.ai / Zhipu AI, the lab behind GLM, as the group behind Ox Alpha. OpenAI and Anthropic each sat at 4 percent. Xiaomi’s MiMo team and Alibaba each drew 6 percent. When a betting market gives one answer ten times the weight of the next tier, that’s not a coin flip anymore. That’s a crowd that thinks it already knows.
I review tools for a living, which means I spend most of my time asking boring questions about who made a thing and what happens to the data I feed it. Ox Alpha makes both questions hard to answer, and it’s still one of the more interesting things to show up in my testing queue this year.
What’s actually confirmed
The verified pile is short. A model called Ox Alpha appeared on OpenRouter and OpenCode on August 20, 2026, listed under a stealth slug. It takes text, images, and video. It reads up to a million tokens in one go. It’s free to use right now. Nobody has claimed it.
By August 23, mainstream coverage out of San Francisco was reporting that developers were piling in, largely because free plus capable is a combination that moves fast in dev circles. Developer chatter on X in the days prior was blunt about the quality, with one widely shared post calling the mystery OpenRouter model insanely good. That’s vibes, not benchmarks, but vibes are how most stealth models get discovered in the first place.
The GLM clue
The strongest thread pointing at Zhipu is a capability gap. Ox Alpha handles text, image, and video, and that multimodal range was not present in the publicly released GLM-5.3. Read one way, that’s evidence against a GLM connection. Read the way the market appears to be reading it, it’s the signature of a lab testing the unreleased version of something it has already shipped a trimmed-down sibling of. Stealth launches exist precisely so a lab can watch real developers hammer a model before the marketing team gets involved.
Worth being honest about the limits here: nobody outside the lab knows. A 63 percent market is a well-informed guess, not a disclosure. The broader speculation that Ox Alpha comes from a Chinese lab is exactly that, speculation, and it’s shaped as much by who has been shipping aggressively as by any hard evidence.
Free access is the part I’d read twice
Here’s where my reviewer instincts start twitching. Free frontier-tier access to an anonymous model is not charity. It’s data collection. Somebody is watching what developers ask, what breaks, where the refusals land, and which prompts push the context window to its limit. That’s the trade, and it’s a fair one as long as you know you’re making it.
Practical version of that, for anyone about to wire Ox Alpha into a coding agent this week:
- Assume prompts and completions are logged. With no known operator, you have no published retention policy to point at.
- Keep proprietary source, customer data, credentials, and anything under an NDA out of it. Test on throwaway repos.
- Don’t build production dependencies on a stealth endpoint. Models under temporary slugs get renamed, rate-limited, or pulled without notice.
- Treat the free tier as a preview of pricing you haven’t seen yet. Budget accordingly if you like what you get.
None of that is a reason to skip it. It’s a reason to sandbox it.
What I’d actually test it on
A million-token window plus video input is an unusual pairing, and it suggests use cases that most current toolchains handle badly. Feeding an entire mid-size repository in one request instead of chunking it through a retrieval layer. Handing it a screen recording of a bug reproduction and asking for a diagnosis. Dropping in long design docs alongside the code they describe and asking where the two disagree.
Those are the tasks where a big context window either earns its keep or quietly falls apart in the middle. Long-context models have a habit of losing track of material buried in the middle of a huge prompt, and no spec sheet tells you whether this one does. That’s a hands-on question, and free access is the cheapest possible way to answer it.
Why the anonymity matters more than the model
Stealth launches have become a standard release valve. Ship under a codename, collect real usage, reveal later. What’s changed is the scale of what gets shipped this way. Ox Alpha is being described as frontier-class by the people using it, and it arrived with no model card, no lab, no safety documentation, and no terms anyone can cite.
Developers are adopting it anyway, because it works and it costs nothing. That tells you something about how the tooling space actually makes decisions, and it isn’t flattering. My advice is simple: test it hard, log what you find, and keep your real work out of it until somebody puts their name on the thing.
đź•’ Published: