alibaba-cloud API Pricing

Short answer. Alibaba prices the same model differently by region — Singapore costs about 17% more than Beijing and the global regions for Qwen3.8-Max — and offers the only explicit cache-creation charge in this database. The list price is therefore a function of where you deploy, which makes region a pricing decision, not just a latency one.

The ladder

Model	Input	Output	Out/in	Cache read	Singapore premium
Qwen3.8-Max	$2.00	$6.00	3.00	$0.25 (12.5%)	~17.5% over Beijing
Qwen3.8-Flash	$0.15	$0.47	3.13	$0.016 (10.7%)	~25% over Beijing
Qwen3.7-Plus	$0.40	$1.60	4.00	—	—

The Singapore premium is not uniform: Flash carries a larger regional gap (25%) than Max (17.5%), which means the cheapest tier is also the most sensitive to where you buy it.

Explicit caching has a creation cost

Alibaba charges $2.50 per 1M tokens to create an explicit cache and $0.17 per 1M to read it. The read price is 8.5% of input, in line with competitors; the creation charge is unusual and changes the arithmetic.

A cache pays for itself only after a few reads: at $2.50 creation against $2.00 input, the first reuse already saves money, but a cache that is created and read once costs more than not caching at all. Short-lived contexts should skip explicit caching and rely on the implicit discount.

Batch pricing is tier-dependent

Batch inference is 50% off on Max and unavailable on Flash. That inversion is worth noticing: the model most likely to be used for bulk classification — the cheapest one — is the one that cannot use the batch tier. Bulk work on Flash pays full rate.

Free quota

1M tokens for 90 days on the Singapore region is a genuine evaluation allowance, roughly enough to run a small pilot. It does not extend to batch or to the Beijing region.

Our view

Region is the first pricing decision with Alibaba, and most buyers make it on latency. The 17-25% premium for Singapore buys data residency and lower latency to Southeast Asia; buying it without needing either is an avoidable cost.

The explicit-cache creation fee makes caching a commitment. Providers that charge only for reads let you cache speculatively; Alibaba makes you amortise the creation. Model it before enabling it on short-lived prompts.

No batch tier on Flash is the biggest structural gap here. Competitors discount batch work by half across the board. If your workload is bulk inference on the cheapest model, Alibaba is more expensive than its list price suggests.

The three-tier ladder is shallow. Max, Plus and Flash span 13x on input — narrower than Zhipu’s 9x flagship-to-flash spread when normalised, and without a speed tier in between.