Build2 publishers3 min readPublished
Filling GLM-5.3-Flash's million-token window costs three times its per-task benchmark price
Z.ai's 320-billion-parameter model activates 18 billion per token and ships under MIT, so a buyer can download it and measure for themselves. Every capability figure published so far comes from Z.ai's own launch materials.
The Engineer · Build desk
What happened
- The weights are on Hugging Face under an MIT license, and the model accepts up to a million tokens of context and returns up to 128,000.
- Before the reveal, Z.ai ran the model anonymously as ox-alpha on OpenRouter and OpenCode, where it became the most-used model of that week with its authorship unknown.
- Z.ai published the launch benchmark table itself, and the dev.to explainer that compiled the figures says to read them as a launch snapshot instead of an independent study.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision The dev.to explainer argues this is now the default model to benchmark against before paying frontier prices, and downloadable MIT weights make that a question of who spends the engineer-days.
- constraint A buyer who wants a capability number from outside the vendor has to generate it in-house, because the published scores all trace back to Z.ai's launch.
- cost Sizing a long-context agent from the $0.045-per-task index figure will understate the bill: one full window of input alone runs about three times that at quoted hosted rates.
- capability If Z.ai's account of the ox-alpha traffic holds, an Nvidia cluster is no longer the gate on serving a model of this class at the volume that topped OpenRouter for a week.
A Mixture-of-Experts model splits its weights into expert sub-networks, and a router turns on a subset of them for each token. GLM-5.3-Flash activates 18 billion of 320 billion [2]. That is 5.6 percent of the parameters per token [1]. Z.ai's documentation credits the hybrid sparse-and-linear attention with 3.0x less attention computation and a 4.4x smaller KV cache than the full GLM-5.3 [10], and LM Studio's model notes describe a compression step called IndexPool, applied to the indexer key vectors, as a large part of why a million-token context is economically servable at all [11]. Pretraining is given as 30 trillion multimodal tokens [12].
Now the hosted price. Third-party serving is quoted at roughly $0.15 per million input tokens and $0.50 per million output [6]. A request that fills the input window and the output budget bills $0.15 for the prompt and about $0.064 for the completion, roughly $0.21 in total [2]. The launch materials put the model at 57 on the Artificial Analysis Intelligence Index at about $0.045 per task, discounted [5]. One full window of input costs a little over three times that per-task figure [3]. The dev.to explainer notes that pricing and quota terms change often and calls its own numbers a snapshot [16].
The AutomationBench move, 26.2 to 48.8 in one generation [13], is 22.6 points and an 86 percent relative gain [4]. For that gain to show up in your budget, your agent runs need the step-chain lengths that harness uses and tool surfaces of the same kind, with no human stepping in between steps.
The license is where the procurement step changes. MIT weights on Hugging Face mean the comparison against a frontier API costs engineer-days and hardware [4]. One item in the record is not a published score: Z.ai ran the model anonymously under the codename ox-alpha on OpenRouter and OpenCode, and it became the most-used model of that week before anyone knew who built it [8]. Z.ai also says that traffic was served on a large cluster of domestically built Chinese AI chips with a custom inference engine, not on Nvidia hardware [9]. Z.ai's own positioning puts Flash on the cost-versus-intelligence Pareto frontier [17].
Vercel's changelog lists GLM 5.3 FlashX on its AI Gateway as a high-speed serving option at about 200 tokens per second [18], addressable as zai/glm-5.3-flashx [19]. Vercel says the gateway reflects provider pricing with no markup and charges no platform fee on inference, including on bring-your-own-key requests [20]. At 200 tokens per second, a 128,000-token response takes about 640 seconds [5]. Eleven minutes is a long stream to hold open for a user watching output arrive. The changelog calls FlashX a serving option for Z.ai's multimodal coding model and does not say how it relates to the MIT-licensed Flash weights [18].
What to watch
- An independent AutomationBench and Artificial Analysis run on the downloaded MIT weights would test Z.ai's launch table.
- Whether the ~$0.15 in / $0.50 out third-party rates hold, given that the explainer says quota and pricing terms change often.
- Whether any gateway documents how FlashX relates to the open Flash weights, and whether 200 tokens per second holds under concurrency.