Z.AI Posts GLM-5.3-FlashX: 200 Tokens/Sec, $0.37/$1.25 per Million

Z.AI's documentation now lists GLM-5.3-FlashX as live, running at 200 tokens per second under the model code glm-5.3-flashx, with a 1 million token context window. The listed prices are $0.37 per 1 million input tokens and $1.25 per 1 million output tokens. That is the full extent of what the docs confirm as of September 18, 2026. No benchmark comparisons, uptime guarantees, or regional availability details appear in the listing.

For a team building a customer support bot or a document pipeline, throughput and context length are not abstractions. A 1 million token context means a manager can feed in long contracts or chat histories without chunking them manually. Speed at 200 tokens per second determines whether a live chat feature feels instant or sluggish. These are the two numbers that decide whether an engineering lead swaps a slower model out of a production system this quarter.

The price split matters more than it looks. Output tokens cost more than three times what input tokens cost, a gap that rewards workloads with short prompts and long answers, and penalizes anyone pasting large documents for short summaries. A workflow that dumps a 50-page PDF into the prompt to get a five-sentence summary pays mostly for the FlashX read, not the write. Finance teams tracking API spend should model cost by task shape, not by a blended per-token average, since the mix changes the bill fast.

What the documentation does not say matters too. There is no stated service level agreement, no latency variance disclosed, no information on how the 200 tokens per second figure was measured, whether under load or in isolation. A policy watcher comparing vendor claims across providers has no standardized benchmark to anchor against here, only a vendor-reported number. Anyone citing this figure in a procurement decision should ask for independent throughput testing under realistic concurrent load before writing it into a contract or an internal rollout plan.

The risk sits differently for each reader. A developer choosing a model code for a new build takes on integration risk if the listed speed does not hold under production traffic. A manager approving budget takes on cost risk if output-heavy use cases dominate actual usage patterns. A policy watcher tracking AI infrastructure pricing takes on comparison risk, since $0.37 and $1.25 per million tokens only mean something against a competitor's equivalent figures, which the documentation does not provide.

Before switching a production workload to glm-5.3-flashx, the next step is straightforward: run a representative sample of real prompts through it, measure actual tokens per second under concurrent requests, and total the input-output cost split against current monthly volume. The documented numbers set a ceiling on expectations. Whether they hold in practice is untested outside Z.AI's own listing, dated September 18, 2026.