The 1/40 figure is a per-token ratio measured against a list price, and the dev.to post carrying it says so before anyone else can [4][15]. Convert it into the unit that decides routing, which is cost per completed task. Apply the stated ratio to the post's own worked session and the same million-in, million-out job against Opus 4.8 list money runs about $60 [1]. A router can therefore attempt the cheap model forty times, or attempt it once and escalate thirty-nine failures, before it has spent what one flagship pass costs [2]. Retry tolerance is what the price gap actually buys.
Two things eat that allowance. The first is a discount you already hold: the post's caveat is explicit that the ratio is against list, so a negotiated Opus rate shrinks it before you write any code [15]. The second is the checker. Escalation only pays if something cheap decides the cheap answer was wrong. If the trigger is a failing test or a rejected schema, you keep most of the headroom. If the trigger is a person reading the output, review labour swamps both token bills and the ratio stops being the number that matters [16].
The price is possible because of the A18B in the name. Around 18B of 320B parameters are active per token, roughly 5.6 percent of the network doing arithmetic on any forward pass [1][5].
The post files this in the same commodity band as DeepSeek V4 Flash [10]. Measured against the rates in GLM's own worked example, that band contains a factor of four: input runs about 2.1 times DeepSeek's, output about 4.3 times [4]. Zhipu's claim that the overall bill still lands below an adjusted DeepSeek V4-Flash is the vendor's own [11], and the price table does not demonstrate it. For it to hold on your traffic, GLM would have to finish tasks in meaningfully fewer output tokens, because output is where the entire gap sits [3].
The tie with Opus is one composite score on one index, 57 against DeepSeek V4 Pro's 53 [2][3]. An aggregate tie is compatible with a large deficit in any single category, and the index does not exercise your tool schema, your prompt formatting, or your latency budget. For the number to transfer, your task mix has to resemble the index's eval mix, and your prompts have to survive a different tokenizer and a different tool-call dialect. The free tier is 200 requests a day [14], enough to sample your prompt set, not enough to load-test it.
One detail to read before planning around the promotion: in the post's arithmetic the discount window takes output from $1.20 to $0.60 per million and leaves input at $0.30 [5][6]. Input-heavy work gets nothing from it. Ten million input tokens is $3.00 either way [7][7], and repository re-reads and long carried agent state are exactly that shape.
The MIT licence promises an exit, though the hardware bill described below keeps that exit out of reach for most teams right now. At two bytes per parameter, 320B weights is 640GB of memory before any KV cache for a 1.04M-token context [1][12][6]. Eight 80GB cards hold the weights exactly and nothing else. The API price is the price, and the licence is insurance against that price changing.
In my context the tiering is worth building: send high-volume classification and document processing through an automatic validator, track the escalation rate, and let the flagship handle whatever fails [17]. I would not route single-shot work with no automatic check through the cheap tier, because there the 40x is a per-token figure and the real cost is the person who reads the answer. The budget under pressure is any budget that assumed all traffic goes through the flagship rate.
