In six days, an anonymous model went from "the internet's favorite guessing game" to an official open-weight release — and Zhipu's playbook deserves a full industry post-mortem.
On August 20, a model called Ox Alpha quietly appeared on OpenRouter: the provider was listed only as "Stealth", it was free with near-unlimited usage, offered a 1M-token context window, and was positioned for coding and agentic production workloads. OpenCode revealed the provider had capacity for 100 trillion tokens per day. Within two days, the developer community started collectively cracking the case.
How the community solved it early
On August 22, researchers triggered server errors with deliberately malformed requests and obtained a Java stack trace whose internal class names mapped directly to Zhipu's documented API routes; the error-code dialect and a 30/30 tokenizer probe run all pointed to GLM-5.3. The most persuasive control: the same GLM weights served on DeepInfra produced a completely different error format — meaning the signature belonged to the API operator, not just the model. Along the way, Stripe CEO Patrick Collison publicly called it "very impressive."
The name itself became a clue: Chinese netizens linked "Ox" to a low-budget animated film that went viral this summer, reverse-engineering a Chinese origin; and two earlier OpenRouter stealth models — Hunter Alpha and Healer Alpha — were both eventually confirmed as Xiaomi's MiMo, validating "stealth launch" as a proven pattern.
August 26: the reveal
Zhipu first confirmed to Bloomberg that morning that Ox Alpha came from its GLM series, with weights promised that night; at 7:42 PM it was officially named GLM-5.3-Flash. Bloomberg reported it had already hit #1 on OpenRouter's usage leaderboard, more than doubling DeepSeek's usage — the marketplace's biggest single-model launch to date.
The specs: a 320B total / 18B active MoE architecture, natively multimodal (text, image, video), 1M context, weights on Hugging Face under the MIT license. API pricing is $0.15/$0.50 per million tokens (input/output), with cached input at $0.03.
On benchmarks, the biggest gains over GLM-5.2 are in agentic capability: DeepSWE 63.4 (previous generation: 46.2), AutomationBench 48.8 (previous: 26.2), and it passes Claude Opus 4.8 on GDPVal-AA v2. But it is not a blanket frontier win — GPT-5.6 Terra still leads on DeepSWE (69.6) and Terminal Bench (87.4).
Why the stealth launch was smart
Zhipu's launch materials are explicit about the strategy: hang the model anonymously on OpenCode/OpenRouter to collect real-world feedback, let it become "the most popular model of the week," then step out under the GLM brand. Stress-testing a frontier checkpoint against production traffic for free, then open-sourcing with proof of demand, carries far less risk and marketing cost than a traditional launch event. One more detail worth watching: the company says the entire preview was served on Chinese AI chips, with the inference stack tuned for a 3× end-to-end improvement — a single-party claim, but if it holds, it means MoE serving at this scale does not inherently require H100 clusters.
So: the stealth free lunch is over, but the era of open-weight agent models at flash-tier pricing is just beginning. When choosing, remember one thing — run your own evals, because a launch benchmark table always shows the cut that flatters the publisher most.
Sources: Business Insider, explainx.ai launch coverage