What is the median price of a Chinese chat LLM API in 2026?

Across the tracked chat models the median paid input price is about $0.4 per 1M tokens, with the cheapest at $0.043 and several flash models free and unlimited.

Are these prices marked up by the aggregator?

No. The figures are raw upstream provider list rates passed through with 0% platform fee (vs OpenRouter ~5.5% markup).

Where does the data come from?

First-party: auto-generated daily from the live per-model pages on aiapi-pro.com. Machine-readable feed at /china-ai-leaderboard.json; interactive view at /china-ai-leaderboard.html.

State of Chinese AI Model Pricing 2026: What 100+ Models Actually Cost

First-party data · generated 2026-10-09 from the live leaderboard

Most "AI pricing" content recycles stale vendor tables. This analysis is computed directly from our live leaderboard of 112 Chinese AI models — real pass-through prices, refreshed daily — so every figure below matches what developers actually pay right now.

Headline numbers

112 models tracked across 6 categories (chat, video, image, audio, 3D, embedding).
5 models are free and unlimited — no token cap, no card: GLM-4.7-Flash, GLM-4.6V-Flash, GLM-4.1V-Thinking-Flash, CogView3-Flash, CogVideoX-Flash.
Chat/LLM paid input prices span $0.043 to $2.9 per 1M tokens (median $0.4) — a >10x spread between budget flash and flagship reasoning models.
0% platform fee: these are raw provider list rates, not marketplace-marked-up prices.

Per-category price distribution

Category	Models	Free	Cheapest	Median	Highest	Cheapest model
Chat / LLM	52	2	$0.043	$0.4	$2.9	GLM-4.7-Flash
Video	11	1	$0.067	$1.001	$1.54	CogVideoX-Flash
Image	23	2	$0.0215	$0.034	$1.154	GLM-4.1V-Thinking-Flash
Audio	13	0	$0.013	$1.54	$1.54	Qwen TTS
3D	7	0	$0.185	$0.323	$0.554	Hunyuan 3D Motion
Embedding	6	0	$0.0769	$0.08615	$0.11	Kinfra VL Embedding 2B

Why the spread matters

The gap between cheapest and median is where most wasted spend hides: for prototyping and high-volume agents, a flash model at the low end often suffices, while flagship reasoning models justify their premium only for hard tasks. Routing by task difficulty — not habit — is the single biggest cost lever.

Billing mechanics behind the numbers

Chat/embedding: per 1M tokens (input/output priced separately on model pages).
Video: per finished second or per upstream token; failed generations refund automatically.
Image: per image; 3D: flat per generation, refunded in full on failure.
Audio: token-billed at the same rate as image/video.

Use the data

The full dataset is machine-readable at /china-ai-leaderboard.json and visualized at /china-ai-leaderboard.html. Comparison sites, agents and cost calculators are welcome to consume it; it updates daily.

See the live leaderboard or start free with $2 credit.