Picture this: an AI model appears on the internet one Thursday morning with no company logo, no white paper, and no name attached. It is free. It has a context window big enough to swallow an entire codebase. And for a week, nobody on Earth is willing to claim it.
That was Ox Alpha, the “stealth model” that showed up on OpenRouter as stealth/ox-alpha on August 20. It was described as a reasoning model built for coding, long-running agent work, and production workloads. The listing credited it to “a third-party provider who has chosen to remain anonymous during this preview.”
Today, the guessing game is over. As reported by Bloomberg, the model is a new entry in the GLM family from Chinese lab Z.AI, formerly Zhipu AI, and the weights are being released tonight.
First, the specs that broke the internet
Let us talk numbers, because this is where the story stops being about gossip. Per the official OpenRouter listing, Ox Alpha shipped with a context window just over one million tokens, a maximum output of 131,072 tokens, and native text, image, and video input. That makes it the first natively omnimodal model in OpenRouter’s anonymous Stealth series.
It was offered free for roughly a week at zero dollars per million tokens. Zero data retention. A knowledge cutoff stretching to November 2025. For developers, that reads like a gift horse.
The benchmark that made it go viral
Within a day, developer Ben Davis ran an informal 10-task pass on the DeepSWE benchmark and clocked Ox Alpha at 80% first-pass accuracy. Claude Fable 5 managed 65%. GPT-5.6 Sol managed 52%. TechCrunch covered the frenzy as it broke.
That single data point was enough. It knocked DeepSeek off a 56-day run atop the OpenCode leaderboard and sent half of AI Twitter into a frenzy. An anonymous model, cheaper than free, apparently out-coding the big names? The internet does not handle that well.
The internet archaeology phase
What followed was a 72-hour guessing game across Reddit and Hacker News. Two theories took shape early: Xiaomi, whose MiMo team had run an identical stealth-launch playbook with the alias Hunter Alpha, and Z.AI, based on a tokenizer fingerprinting test that Economic Times broke down for a general audience.
The fingerprint evidence piled up fast. Ox Alpha’s raw token counts matched GLM’s tokenizer on 11 out of 11 probes. On a digit test it burned 29 tokens, an exact GLM match, where DeepSeek used 98. It even stumbled on the same “dirty token” that has historically tripped up Qwen and GLM-family models. The Next Web linked the same fingerprint trail to censorship concentrated on seven topics.
Then came the chaos. An anonymous GitHub gist showed Ox Alpha solving olympiad-level math. Two prominent AI-safety researchers posted cryptic tweets reading “we are not just facing a new model,” then deleted them. OpenAI and Google declined to comment. Somewhere in there, someone called it an AI UFO.
Today’s reveal
Bloomberg settled it. Ox Alpha is a new iteration of the GLM series, the flagship release from Z.AI. The lab plans to release the model’s weights tonight, making it open for anyone to run, under the same MIT license Z.AI has used since GLM 5.
It fits a pattern the industry already knew. Past Stealth anonymous models were eventually revealed as GLM-5 or MiMo releases, an “anonymous warm-up to collect real usage data” playbook with precedent. Z.AI founder Tang Jie had also teased a multimodal mode three months earlier.
What it means for everyone else
The uncomfortable part for Western labs is the math. Z.AI prices its GLM 5.3 at $1.40 per million input tokens and $4.40 per million output tokens, with an 81% cache discount. A model that looks competitive with Fable 5 and GPT-5.6 Sol on coding, released openly at a fraction of the cost, is exactly the pressure point closed labs have spent two years avoiding.
There is a sobering footnote too. Researchers say Ox Alpha can reason across an entire exploitation chain rather than flagging isolated bugs, a capability Z.AI says it did not deliberately optimize for and frames as an emergent property of scaling. The moment the weights go public, that capability spreads.
The bottom line
For about a week, a frontier-adjacent model was free on a public API with no name tag, and the internet could not decide if it was a miracle or a performance piece. It turned out to be a Chinese lab’s quiet field test.
Ox Alpha never needed a name to prove itself. It just needed one to go to market. We see you, Z.AI.
