Chinese AI Model Pricing Tracker (September 2026)

How to read the table

The provider families in this snapshot are DeepSeek, Doubao, ERNIE, GLM, Kimi, MiMo, MiniMax, Qwen, StepFun. The table is catalog data, not a claim that every capability or route has been tested for every workload.

Dated base rates

Forecast a workload

Start with a token mix, not a model label. For 50,000 uncached input tokens and 2,000 output tokens, multiply each token count by the matching per-million rate. Add cached input only when the ledger reports a cache hit.

input_cost = input_tokens / 1_000_000 * input_rate
cache_cost = cached_input_tokens / 1_000_000 * cache_rate
output_cost = output_tokens / 1_000_000 * output_rate
base_cost = input_cost + cache_cost + output_cost

Run one controlled request after choosing a route. Compare the model ID, token fields, account group, and charged quota in the request ledger with your estimate.

What the table does not prove

Price does not establish quality, context capacity, tool-call support, or route availability. Check the model catalog and current status page, then run a representative test set. Some capability fields in the internal catalog are intentionally marked unknown until they have direct evidence.

Use the JSON endpoint for automation. Use this page when a person needs to scan the catalog, compare token categories, and retain the effective date beside a forecast.