Z.ai Confirms It Built Ox Alpha, Reveals It as GLM-5.3-Flash

A mysterious model called Ox Alpha appeared without warning on OpenRouter and OpenCode in late August, climbing benchmarks and leaderboards against the best available models with no lab willing to claim it. Over the following days, that silence turned into one of the AI industry's biggest guessing games: our earlier coverage of the stealth launch tracked five rival theories about who had built it — Z.ai's GLM family, Microsoft, Google, Cursor, and ByteDance — based mostly on tokenizer fingerprinting, since none of the named labs would confirm or deny anything. Tokenizer probes pointed most consistently at Z.ai, matching the behavior of its GLM models on eleven separate tests where no other lab's models matched on more than four, but the sheer scale of free usage Ox Alpha was absorbing left even that leading theory in doubt.

That mystery is now resolved. Z.ai has confirmed, according to a Bloomberg report cited by TechCrunch, that Ox Alpha is the newest member of its GLM series. The company said it plans to release Ox Alpha's weights on Wednesday, after which outside developers will be able to build on top of it directly. Z.ai describes the model as built for coding, extended agentic work and production use, aimed at long-running software-engineering tasks and workflows that mix text with images.

Behind the mask: GLM-5.3-Flash

Under its formal name, Ox Alpha is GLM-5.3-Flash — Z.ai's first natively multimodal release in the GLM-5 series, according to the German outlet the-decoder.de. It carries 320 billion total parameters, of which only 18 billion are active on any given task, a mixture-of-experts design meant to hold down inference costs. The model ships under an MIT license, with its weights posted on Hugging Face, and supports a context window of one million tokens.

On Artificial Analysis's Intelligence Index, GLM-5.3-Flash scores 57 points at maximum reasoning effort — three points behind the larger GLM-5.3's 60, and roughly level with GPT-5.6 Terra and Muse Spark 1.2. The gap that matters more is price: Artificial Analysis puts the cost per task in its index at $0.09, against $0.68 for GLM-5.3, about 7.5 times cheaper, which the firm says places the model on the Pareto frontier of intelligence versus cost. Through Z.ai's own API, GLM-5.3-Flash costs $0.15 per million input tokens and $0.50 per million output tokens, roughly a tenth of GLM-5.3's price. On the agentic GDPval-AA v2 benchmark, it scores an Elo of about 1770, matching GLM-5.3 and Grok 4.6 and trailing only Claude Opus 5. The trade-off is efficiency: Artificial Analysis found that around 90% of the model's output tokens go toward reasoning rather than the final answer, which can make responses slower even at a lower price per token.

• 320 billion total parameters, 18 billion active per task (mixture of experts)
• Intelligence Index: 57 points, versus 60 for GLM-5.3 — at about 7.5 times lower cost per task ($0.09 vs. $0.68)
• API pricing: $0.15 per million input tokens, $0.50 per million output tokens — roughly a tenth of GLM-5.3
• GDPval-AA v2 agentic score: Elo ≈ 1770, behind only Claude Opus 5
• Claimed 100 trillion tokens served per day, entirely on Chinese AI chips

Testing Nvidia's 'CUDA moat'

The more striking claim concerns infrastructure, not just the model. Before its named release, Z.ai tested GLM-5.3-Flash anonymously as 'ox-alpha' on OpenCode and OpenRouter, where it became the most-used model of the week. According to Z.ai, all of that traffic ran on Chinese AI chips instead of Nvidia hardware. SemiAnalysis reported the setup delivered 100 trillion tokens per day, a scale previously thought achievable only by frontier labs, and noted that Z.ai claims hardware efficiency and per-token cost comparable to mainstream Nvidia GPUs. SemiAnalysis frames the result as another test of Nvidia's so-called CUDA moat — the proprietary software layer, built up over nearly two decades, that sits between AI frameworks and Nvidia's chips and that most AI software is tuned specifically around. Switching to a different chip normally means reprogramming compute operations and memory access from scratch, work that has historically left Nvidia's rivals with chips that were fast on paper but poorly used in practice. Z.ai says it got around that by building its own serving software on the open-source SGLang framework, splitting inference into independently scalable stages; the team says that tripled throughput on the same hardware compared with its first attempt, with an agent built on GLM-5.3 assisting the optimization work.

Anonymous 'stealth' launches like this one have become something of a playbook, particularly for Chinese labs: releasing a model without a name attached lets it get judged purely on leaderboard results and real developer usage, free of the assumptions that come with a known brand, before the lab commits to a formal launch. It doubles as a way to test a brand-new infrastructure stack at production scale — in this case, a full Nvidia-free serving pipeline — before staking the company's name on the outcome.

The no-Nvidia claim lands against the backdrop of years of U.S. export restrictions that have limited Chinese labs' access to the most advanced Nvidia AI chips, pushing them toward domestic alternatives and custom software to make those chips competitive. If Z.ai's efficiency numbers hold up under independent scrutiny, they weaken the assumption that Nvidia's CUDA ecosystem alone guarantees it a lasting edge over labs building around export restrictions rather than through them — with real implications for how much pricing power Nvidia can keep in a market it has long dominated.

For companies weighing whether to route production workloads to a cheaper Chinese model instead of a Western frontier one, GLM-5.3-Flash adds a concrete data point: benchmark performance close to a full-size flagship, at roughly a tenth of the API price, under a license that allows self-hosting. The trade-offs are real, too — heavier reasoning-token usage that can slow responses, a lab based in China with its own data-handling and export-compliance questions, and specs that, like Ox Alpha's stealth run itself, went unverified by independent testers until Z.ai chose to reveal them. Weighing cost savings against those governance questions, rather than against benchmark scores alone, is likely to become a standard step in vendor evaluation as more low-cost, high-performance releases arrive from Chinese labs.