AUX Alpha (GLM 5.3 Flash): What Happened

From anonymous model to viral hit

• AUX Alpha appeared anonymously on OpenRouter and served 42 trillion tokens in six days, becoming its most-used model and drawing roughly a third of the platform’s weekly traffic at its peak.
• It offered a million-token context window, video input, strong coding ability, and was initially free.
• Its popularity fueled speculation about its creator. Clues included API errors associated with Zhipu’s Z.ai and a tokenizer fingerprint resembling the company’s GLM models.
• Despite warnings that prompts might be retained, some developers used it with proprietary code and routed Claude Code through it to avoid rate limits.
• Viral claims about its benchmark results were exaggerated: its actual score on Deep SWE was reported to be closer to 58%, not the 80% shown in a widely shared screenshot.

The reveal and pricing

• On August 26, Zhipu announced that AUX Alpha was GLM 5.3 Flash, a natively multimodal mixture-of-experts model with 320 billion total parameters.
• Its weights, released under the MIT license, appeared on Hugging Face the same day.
• Zhipu said the stealth run used 100,000 Chinese-made chips.
• The free anonymous endpoint ended, but the model remained inexpensive: $0.15 per million input tokens and $0.50 per million output tokens, with a temporary 50% discount through September 9.

Testing it on a legacy app

The presenter tried the model on Horse 2, a decade-old AngularJS app, asking it to modernize the project while preserving its humor and style.

• It identified the outdated stack, but spent a long time explaining its plan.
• It built a new design using plain CSS, including a purple-sloth background, gradient buttons, and animated video cards.
• It fixed a mobile overflow issue with the familiar min-width: 0 CSS workaround.
• It preserved existing jokes and added new humorous notifications.
• Rather than adopting a modern framework, it rewrote the app in HTML, CSS, and vanilla JavaScript.
• For a horse video, it used FFmpeg to extract frames, analyzed the footage, and generated relevant title, description, and comments.

Takeaways

• The model’s capabilities appeared credible in hands-on use, including video understanding and coding.
• Its main drawbacks were slow responses, excessive narration, and occasional doom loops.
• The video also notes that developers should take privacy warnings seriously before sending sensitive code or prompts to third-party services.

Sponsor: Exa

Exa is presented as a search tool for AI agents, offering web results plus research papers, financial data, and information on people and companies. The presenter used it to retrieve SEC filings and return structured JSON, with source citations.