The week’s most-used open model finally got a name and a price, and both landed tonight. The anonymous “Ox Alpha” that topped OpenRouter’s usage charts for roughly a week — free, no card, no lab, processing tens of trillions of tokens — was confirmed on Wednesday to be Z.ai’s (Zhipu’s) latest GLM, and the company then revealed it as GLM-5.3-Flash: a 320-billion-parameter, 18-billion-active mixture of experts, native multimodal (text/image/video in), 1M-token context, with weights released the same evening under an MIT license. The rate card is $0.15 and $0.50 per million input and output tokens — about a tenth of the $1.40/$4.40 Z.ai charges for GLM-5.3 itself, and well under the $0.25/$0.90 median for its price tier on Artificial Analysis. The open layer just priced its own flash model, and it shipped a license, not a sales pitch.
The stealth reveal is the pricing-power thesis with a signature
What happened. Ox Alpha appeared on OpenRouter as a free “stealth” entry around August 20 and out-used every other hosted model for about a week. Bloomberg reported Wednesday that Z.ai confirmed the model as a new GLM-series iteration and would release its weights the same night; Z.ai then identified it as GLM-5.3-Flash and pushed the weights — MIT-licensed on Hugging Face on day one, with a week of free API access and a 50% launch discount live on OpenRouter through September 9. The hardware story: a new hybrid sparse-plus-linear attention architecture (NoPE-MLA with KDA linear attention per the GB10 deployment write-ups), native image/video input, 1M-token context, day-zero support in vLLM and SGLang, and an independent Artificial Analysis intelligence index of 57 against a ~18 median in its price tier. Qwen’s open-weights preview this morning carried a community license; GLM-5.3-Flash shipped permissive MIT out of the gate.
Why it matters. This is the open-weights pricing power argument’s loudest data point yet: Z.ai undercut its own flagship coding model nine-to-one in under two weeks, while beating it to a permissive license — GLM-5.3 launched August 14 with weights promised “in two weeks,” and the Flash edition’s weights were out the same day it got a name. Demand was proven silently, by topping the router’s usage chart before anyone knew who made it, not by launch-day hype. For the operator the move is to reroute agentic and coding traffic that tolerates open weights to this tier now; the reality check is the 320B total footprint — on two DGX Sparks the first published GB10 run needed a kernel workaround for the new attention layout, and hit ~25-30 tok/s at a 32K context cap, text-only, in that configuration. Cheap frontier flash is a cloud-tier bet; it is not yet a desktop one.
What I’m watching
Whether the promised weights actually land complete and under the MIT license as stated, and whether independent re-runs keep the vendor coding claims honest — that is the verify-later reflex the next 11:00 column should score. And whether either big US lab answers a 9x-cheaper open flash with anything but silence, because that silence is now the price signal.
Source: openrouter.ai, bloomberg.com, techcrunch.com, artificialanalysis.ai, forums.developer.nvidia.com