Eighteen billion active parameters out of 320 billion is 5.6 percent of the model doing work on any given token [3][1], and that ratio is where a price claim of this shape comes from. Compute per token tracks the active count; the checkpoint a team has to host tracks the total, and the other 302 billion parameters still have to sit somewhere [2]. The card credits the long-context saving to a hybrid of sparse and linear attention, new to the GLM series, plus Manifold-Constrained Hyper-Connections for scaling efficiency [6][7]. It gives no memory footprint, no hardware target and no quantisation guidance [24], which is the first thing a team weighing self-hosting needs.
A skeptic would say a tenth is a tenth, whoever is being paid. The difficulty is the comparator: the tenth is stated against GLM-5.2 [4], and the README names no per-token rate for either model [22]. Moving it into a budget requires the published rate for GLM-5.2 on the Z.ai platform [20], then a second comparison against the incumbent invoice a team actually holds. Until both sit on the same page, the figure describes a vendor's pricing decision about its own catalogue rather than anyone's unit economics.
The second thing to price is tokens per task. Benchmark reproduction requires reasoning_effort at its default max, the top of three levels [10], and the evaluation notes show what that regime consumes: 163,840 tokens of maximum generation length on HLE with tools inside a 300,000-token context [12], a 400K context and a six-hour timeout on DeepSWE [14], 64k new tokens under a 1M context on NL2Repo [13]. Turning the effort down to low or high to save money moves a deployment off the configuration in which the reported quality was measured [10], and the card's text carries those configuration notes without the score tables themselves [23].
Two of the named evaluations come from outside the vendor. GDPval-AA v2 was run by Artificial Analysis [17], and Toolathlon Verified results were obtained through the official evaluation service as pass@1 averaged over three runs [16]. Terminal-Bench 2.1 was run inside Claude Code 2.1.207 [15], a harness the vendor did not write even though it still reports the number, and HLE was graded by GPT-5.6-luna at medium [12]. NL2Repo added rule-based and model-based checks against unauthorized pip or curl operations [13], which says the team expects agents to look for shortcuts. That is a reason to take the harness design seriously and a reason to run the workload in-house before signing anything.
For this quarter, the licence is the part of this release a buyer can act on [1]. A 320B checkpoint with documented SGLang, vLLM and Transformers support [9] is a credible alternative to name in a renewal conversation without ever being deployed, and MIT terms mean that option does not lapse when the next model lands. The case for actually replacing an incumbent needs third-party scores the card does not yet contain [23], and that is next quarter's evidence, not this one's.
