Z.ai Ox Alpha model tops charts as a stealth release

Z.ai Ox Alpha model has done something unusual in artificial intelligence, sweeping to the top of global usage charts before anyone even knew who made it. The mysterious model that racked up hundreds of millions of calls in a single week has now been confirmed as the work of China's Z.ai , the Beijing company formerly known as Zhipu. Z.ai revealed that Ox Alpha is a test version of its GLM-5.3-Flash model, deployed anonymously as a stealth release to gather real-world usage data and honest developer feedback before its official launch, and the secret behind its viral rise is a potent combination of high performance, near-zero cost, and a deliberate cloak of anonymity. The stealth strategy at the heart of the story is a clever and increasingly common tactic. Ox Alpha appeared on the AI platforms OpenRouter and OpenCode on 20 August with no company logo and no country of origin, attributed only to an anonymous third-party provider, and it was offered free with no rate limits for a week. By hiding its ownership, Z.ai could test its frontier product in the wild without the weight of expectation or brand bias, letting the model be judged purely on merit. It worked spectacularly. Within a day, an informal developer benchmark had Ox Alpha outperforming rivals on coding tasks, and it went viral, ending DeepSeek's 56-day reign atop the OpenCode leaderboard. Over the week it was called more than 340 million times and processed over 27 trillion tokens, and detective work by developers, who matched its tokenizer fingerprint to Z.ai 's GLM series on all 11 probes, unmasked it well before the company confirmed the truth. The competitive significance is what makes this more than a marketing stunt, and it lies in the pricing. Z.ai intends to charge around 0.15 dollars per million input tokens and 0.50 dollars per million output tokens, placing Ox Alpha, now GLM-5.3-Flash, firmly in the class of low-cost, high-efficiency models pulling users away from the premium offerings of American labs like Anthropic and OpenAI. The model is described as a reasoning system built for coding, sustained agentic work and production workloads, handling text, images and video with a context window over a million tokens, and Z.ai will release its open weights so developers can build on it freely. A Chinese lab landing a model that appears competitive with frontier Western systems on coding and agentic benchmarks, at a tenth to a twentieth of the price, changes the calculus for every company currently selling frontier access as a premium product. The regional dimension matters for the Gulf specifically. Cheap, capable, open-weight Chinese models are enormously attractive to markets like Saudi Arabia and the UAE, which are building sovereign AI capabilities and want to control their own infrastructure without paying premium Western prices or depending entirely on US providers. Open weights can be downloaded, run locally and adapted, including for Arabic-language use, which aligns neatly with the region's data-sovereignty priorities. Notably, Z.ai disclosed that all of Ox Alpha's inference ran on Chinese-made AI chips, a demonstration of China's growing self-reliance that will resonate with Gulf states weighing how to diversify their own AI supply chains amid US export controls. The honest caveats are worth noting. Benchmark performance, especially from informal early tests, does not always translate to real-world reliability, open-weight models raise their own safety and governance questions, and the pricing war compressing AI margins may not be sustainable for anyone. But the Z.ai Ox Alpha model is a genuine milestone, showing that Chinese labs can now match frontier capability at a fraction of the cost, and that with four of the world's top-10 models now Chinese, the open-weight race is reshaping the global AI landscape in ways the Gulf and everyone else will have to reckon with.