China AI model radar: what's actually live this week

Every Friday we will publish the state of the catalogue: what is live, what moved on price, and what we are still testing. No rankings we cannot reproduce, no benchmark scores we have not run ourselves.

The catalogue, this week

The platform currently serves 90-plus models across the major Chinese families — DeepSeek, Qwen, GLM, Kimi, Hunyuan, ByteDance's Doubao and Seedance lines, MiniMax, Kling, and others. The full list with live pricing is in the Model Square; it changes more often than this post will.

What is structurally notable:

• Text and reasoning are priced in single-digit dollars per million tokens for most models, with cached input often an order of magnitude cheaper.
• Video generation is billed per call or per task, and dominates cost in any workload that touches it. A single short clip can cost more than a day of chat traffic.
• Image generation sits between the two: cheap per call, but easy to run at volume.

Price positioning, without stale numbers

We deliberately do not print price tables in blog posts. Upstream pricing changes, discounts rotate, and a post with yesterday's number is worse than no post. What we will do instead:

1. Describe the shape of the market — which tier each family competes in.
2. Link to the exact model page, which always shows the current published price.
3. Flag moves in this post when they happen.

This week's shape: DeepSeek and the Qwen hosted tiers anchor the value end for text. GLM competes on agentic coding rather than price. Kimi prices long context at a premium, which is rational — it is selling a capability, not tokens. ByteDance and MiniMax compete on throughput and per-call generation.

What we are still measuring

Publishing benchmark numbers is the easy part. Publishing reproducible ones is not. Before we put a score on the benchmark index, we want:

• a pinned model ID and provider
• a fixed prompt set and decoding parameters
• cost per completed task, not just accuracy
• a date and a way to re-run it

Until then the benchmark page shows pricing and capability metadata only, with an explicit "not yet measured" state. We would rather ship a blank column than a number we cannot defend.

Watching this week

• Agentic coding variance. The same task, same model, same temperature: outputs still diverge enough that a single run tells you nothing. If you evaluate, run each task at least three times.
• Cached pricing adoption. Fewer teams than you would expect enable prompt caching in agent loops. It is the cheapest performance win available this month.
• Video cost drift. Per-call video pricing is the least stable part of the catalogue.

How to use this

If you are choosing a model this week: start from the vendor pages, check current pricing on the model, then run ten of your own tasks. If you are building: the local deployment guide is still the fastest path to a working endpoint.

Corrections and re-run results are welcome at research@chinesemodels.ai. If you find a cheaper way to get the same completed-work number, we will publish it.