An anonymous model appeared August 20 and was in production within 24 hours.
On August 20, a model named "stealth/ox-alpha" showed up on OpenRouter. No company name, no launch event, no pricing, just a one-week free preview.
And it showed up writing code that put GPT-5.6 Sol on the floor.
How brutal were the scores
DeepSWE Pass@1 hit 80%, leaving the famous names behind: GPT-5.6 Sol at 52%, Claude Fable 5 at 65%, Zhipu's GLM-5.3 at 62%. Context window 1,048,576 tokens, max output 131K, fully multimodal, text, image, native video understanding all supported.
Function calling, structured JSON output, and agentic workflows shipped in the box. In other words, it is not a chat toy but a production-grade base that can do real work.
What is strange is the adoption speed
Anonymous models usually die within a week. This one is different: within 24 hours of launch, Nous Research's Hermes Agent and the Zed code editor wired it into production.
An unnamed model picked up by two well-known tools on day one means either it is absurdly strong, or the people and resources behind it are anything but ordinary. A normal open-source model takes weeks or months to reach mainstream tool integration.
The community is nearly certain it is Zhipu
Independent researcher Ben Davis ran a technical fingerprint comparison and concluded with 99% certainty that OX Alpha belongs to Zhipu's unreleased GLM-5.x multimodal flagship. Evidence includes an identical video-encoder token-consumption signature.
If it really is Zhipu, this is the first time a Chinese lab grabbed a global board slot by "anonymous airdrop" rather than an official launch, letting the community and developers vote with their feet first. The tactic is more interesting than the model: manufacture mystery for attention, keep users with strength.
Read it inside the August model flood
OX Alpha is not isolated. The whole of August was too dense to test: Alibaba open-sourced Qwen3.8-Max at 2.4 trillion parameters, DeepSeek V4 Pro went GA, Meta's Muse Spark returned to open weights, Tencent shipped Hy3, Zhipu's GLM-5.3 climbed via post-training. Five-plus vendors shipped 11 models in 20 days. OX Alpha is the wildest drop in that flood, too lazy to even give a name.
That itself signals a trend: the capability gap between frontier models is closing, and who shows up first no longer decides the win. The way you show up, and the word of mouth, are starting to count.
For builders, the takeaway is not which lab wins this round. It is that the barrier to a credible frontier release keeps dropping, and the launch playbook is being rewritten in public. A model with no name out-scored named flagships within a day. Reputation lag is now a real thing in this market, and incumbents can no longer assume their brand alone buys them the top slot.
A few cold showers
Anonymous plus free preview means no independent replication and no terms-of-service backstop. You can use it now and it feels good, but after a week it may reprice, throttle, or simply vanish. Do not run core business on it, and do not wire it into production architecture as a long-term dependency.
Also, leading on one coding axis is not leading overall. GPT-5.6 and Fable 5 remain all-rounders across more dimensions. OX Alpha reads more like a precision deterrent shot, forcing rivals to re-examine the strength of China's open-source camp, not a full overtake.
Why this matters to you
If you write code, treat OX Alpha as a week-long free toy, test its coding and video understanding, but do not depend on it. The real signal worth watching: Chinese models have shifted from "hold a launch" to "airdrop and grab the board," which says both capability and confidence are up. For your model choice, stop staring only at the US labs. Zhipu, Qwen, DeepSeek are increasingly competitive, and usually cheaper. One more option is always good news. When its true identity surfaces, whichever lab it is, go re-run the scoreboard.