The stealth AI model that recently dominated OpenRouter under the name "ox-alpha" is actually GLM-5.3-Flash, a 320 billion-parameter system from Chinese lab Z.ai that matches leading American models at a fraction of the price. On Wednesday, Z.ai revealed the model's identity and released its weights on Hugging Face under the MIT license, according to a report from The New Stack. The model was purpose-built for ultra-low-cost inference and now runs on multiple platforms at $0.075 per million input tokens and $0.25 per million output tokens—prices that reflect a 50% discount.
GLM-5.3-Flash scores 57 points on the Artificial Analysis Intelligence Index, placing it alongside GPT-5.6 Terra, Google's Gemini 3.7 Flash, Meta's Muse Spark 1.2, and Qwen 3.8 2.4T A95B, the report states. While early excitement compared the model to Anthropic's Claude Fable 5, benchmarks don't support that claim—but the system can largely match Claude Opus 4.8 and OpenAI's GPT-5.6 Terra when operating in maximum-effort reasoning mode. The model handles multimodal inputs including images, videos, and files, with training optimized for visual tasks like presentation and website creation, plus standard knowledge work involving documents, spreadsheets, and dashboards. One trade-off: the model is verbose and consumes significant tokens, though low inference costs offset this issue, according to the report.
Z.ai served 100 trillion free tokens daily during the ox-alpha testing period, all processed on Chinese AI chips rather than NVIDIA hardware. "Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance," the company writes, adding that it reached "hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs." The team built a custom inference engine based on SGLang, using its flagship GLM-5.3 model to power an infrastructure agent that created what the company describes as "a feedback loop in which the model helped optimize the system serving the model itself." The model was trained on a 30 trillion multimodal pre-training corpus using a hybrid architecture combining linear and sparse attention, which cuts compute by 3× and KV cache size by 4.4× compared to the full GLM 5.3 model, the company notes.
For developers, a model matching Opus 4.8 on agent-relevant benchmarks at one-tenth the price is difficult to overlook, the report states. Z.ai stresses that the launch shows Chinese chips can support frontier-model inference efficiently and economically at scale—a claim backed by the company's ability to handle massive serving volumes that few infrastructure providers could match. The broader question for American labs is what happens now that Chinese companies can serve at this scale and price point, particularly as open-weight models from Alibaba, Deepseek, Moonshot, and Z.ai close the performance gap with U.S. frontier models while remaining available at minimal cost or free for those with suitable hardware. The economics of AI inference may be entering a new phase where geographic chip restrictions matter less than architectural innovation and serving efficiency. Decisions about model selection will increasingly hinge on whether proprietary performance edges justify order-of-magnitude cost premiums.