# How Chinese AI providers bill: five units, four traps (markdown mirror)

Source: https://aiapi-pro.com/blog/chinese-ai-billing-units-2026

Catalog on 2026-10-22: 111 models. Counts: 48 Chat/LLM (46 token-billed, 2 free), 22 Video, 17 Image, 11 Audio, 7 3D, 6 Embedding.

## The five units

| Unit | Used by | Typical price | Covers |
|---|---|---|---|
| Per 1M tokens (in/out) | 48 chat, 6 embedding, some image/video | $0.043-$2.90 /1M in | text in/out (embeddings: input only) |
| Per image | 4 image models | $0.0215-$0.14 | one image at default settings |
| Per 5-second clip | 18 video models (16 estimated) | $0.054-$0.70 | 5s at minimum resolution |
| Per character / per token (speech) | TTS models | $0.0145/1K chars (Qwen) or $1.5385/1M tokens (MiniMax) | audio output length |
| Per generation | 7 image, 7 3D, 3 audio | $0.185-$0.323 | one render/asset |

## Four traps

1. **"Per 5-second clip" means lowest resolution.** Seedance 2.5 at $0.70/clip is the floor: minimum resolution, 5 seconds. Higher res/longer/premium modes cost proportionally more.
2. **Some video models are token-billed.** MiniMax Hailuo and others bill tokens computed from duration and resolution. We publish an estimated token rate and derive `cost_per_5s_est` so families land in one column.
3. **Speech units do not convert.** Qwen TTS per character; MiniMax speech per token. Characters and tokens are not convertible without a tokenizer assumption.
4. **Embeddings have no output price.** Input-only billing; `price_per_1m_output` is 0 and ranking uses input only.

## How we normalize

| Field | Meaning |
|---|---|
| `price_per_1m_input` / `price_per_1m_output` | text models |
| `cost_per_5s_est` | video models at minimum resolution |
| `price_per_image` | image models at default settings |
| `normalized_unit` | which unit the row was normalized into (`free` for free models) |
| `price_usd` / `unit` | provider's own number and unit, unmodified, for audit |

## Practical rules

1. Match the unit to your workload, not the price list.
2. Assume the headline is a floor (resolution/length/mode multipliers).
3. For token-billed media, measure once with a real request.
4. Keep the raw unit; when providers change units, only the normalized field stays comparable.

## FAQ

**Why can't I compare a per-image price with a per-token price?**
They measure different things. An image price covers one generation regardless of prompt length; a token price covers text volume. Convert both to the cost of one typical job in your workload.

**What does "per 5s clip at lowest resolution" mean?**
The cheapest configuration of that video model: five seconds at minimum resolution. Higher resolution, longer clips or premium modes cost proportionally more.

**Are any video models billed by tokens?**
Yes. Some are billed in tokens computed from duration and resolution rather than a fixed per-clip price. We publish an estimated token rate and convert it to a per-5s-clip figure.

**How is speech billed?**
Two ways: per character (Qwen TTS at $0.0145 per 1K characters) or per token (MiniMax speech at $1.5385 per 1M tokens). They do not convert.