zhipu-ai API Pricing

Short answer. Zhipu prices three tiers of the same family with a 9x spread: GLM-5.3 at $1.40/$4.40, GLM-5.3-Flash at $0.15/$0.50, and GLM-5.3-FlashX at $0.37/$1.25. The older GLM-5.2 sits at exactly the flagship price, which tells you something about how the company wants you to choose.

The ladder

Model	Input	Output	Out/in	Cache	Position
GLM-5.3	$1.40	$4.40	3.14	$0.26 (18.6%)	Flagship
GLM-5.2	$1.40	$4.40	3.14	$0.26	Previous flagship
GLM-5.3-FlashX	$0.37	$1.25	3.38	$0.075	Speed tier
GLM-5.3-Flash	$0.15	$0.50	3.33	$0.03 (20%)	Volume tier

The output/input ratios are unusually consistent — 3.1 to 3.4 across all four — which means the tiers differ in absolute price, not in shape. Choosing between them does not change how your workload’s cost scales; it changes the multiplier.

FlashX is a latency purchase

At $0.37 input against Flash’s $0.15, FlashX costs 2.47x for a faster serving tier of the same underlying generation. That is a reasonable premium if you are serving an interactive product where time-to-first-token is visible to users, and a bad one for batch work, where it buys nothing.

The China platform prices tell the same story in CNY: ¥2 / ¥7 against ¥0.8 / ¥2.8 — the same 2.5x relationship, regionally consistent.

The GLM-5.2 problem

GLM-5.2 is priced identically to GLM-5.3. There is no discount for the older model, no sunset price, no incentive to move traffic onto the previous generation. A buyer comparing the two should treat 5.2 as a fallback for compatibility, not as a cheaper option — because it is not one.

Our view

Zhipu’s tiering is the cleanest in the Chinese market to reason about. Four models, one shape, a 9x span from Flash to flagship. You can pick a tier on cost and know exactly what multiple you are accepting.

Flash is the default and the flagship is the exception. At $0.15/$0.50 with a 20% cache rate, Flash covers most workloads; the 9x step to GLM-5.3 should be justified by a measured task failure, not by a benchmark gap.

Paying for FlashX instead of the flagship is often the better upgrade. If the problem is latency rather than capability, 2.47x buys speed; 9.3x buys parameters you may not need.

A same-price older model is a signal about inventory. Zhipu is not trying to clear GLM-5.2 capacity; it is keeping it available and priced at parity. Do not assume a legacy discount that does not exist.