Cheap and capable: Z.ai's GLM-5.3-Flash challenges costly rivals from Chinese chips

The mystery model Ox-alpha, which quickly became the most popular option on OpenRouter in recent days, has revealed itself to be Z.ai's GLM-5.3-Flash—a massive yet lean model built specifically to deliver strong performance at rock-bottom inference costs. On Wednesday, Z.ai unveiled the stealth model and released its weights on Hugging Face under an MIT license.

The model boasts 320 billion parameters but only activates 18 billion at inference time, keeping computational demands minimal. Early adopters can access it via OpenRouter at roughly $0.075 per million input tokens and $0.25 per million output tokens—a 50% discount from standard pricing—putting it roughly one-tenth the cost of leading Western alternatives.

Performance and capabilities#

GLM-5.3-Flash doesn't quite reach the level of Anthropic's flagship Fable 5, despite early buzz suggesting otherwise. Yet it trades blows effectively with Claude Opus 4.8 and OpenAI's GPT-5.6 Terra when both are optimized for demanding reasoning tasks. On the Artificial Analysis Intelligence Index, the model scores 57 points, placing it alongside GPT-5.6 Terra, Google's Gemini 3.7 Flash, and other recent releases.

More notably, it excels at driving AI agents—a test that often correlates better with real-world utility than general benchmarks. The model understands multimodal inputs including images, videos, and files, with Z.ai stressing that its training emphasized visual-intensive tasks like presentation and web design creation, plus document-heavy work with spreadsheets and dashboards.

One tradeoff: the model is verbose and consumes tokens liberally—typical of smaller models requiring extra reasoning steps. At these pricing levels, however, that inefficiency barely registers as a cost concern.

Chinese chips handling global-scale inference#

The standout achievement isn't performance; it's the infrastructure supporting it. Z.ai served 100 trillion free tokens daily—a volume few providers can sustain—entirely on Chinese-made AI chips. The company emphasized this capability in its announcement, stating that it achieved "a 3× improvement in end-to-end serving performance" compared with earlier attempts on the same hardware, matching "hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs."

This result signals that Chinese silicon can support production-scale frontier-model serving economically, a finding that complicates the competitive calculus for Western AI labs.