Every provider publishes a price. Almost none publish it in the same unit. Before you can compare two models — or even one model against another from the same vendor — you have to know which unit the number is in, and what that unit hides.
| Unit | Used by | Typical price | What it covers |
|---|---|---|---|
| Per 1M tokens (input / output) | 48 chat models, 6 embedding models, some image and video models | $0.043 – $2.90 /1M in | Text in, text out (embeddings: input only) |
| Per image | 4 image models | $0.0215 – $0.14 | One generated image at default settings |
| Per 5-second clip | 18 video models (16 estimated) | $0.054 – $0.70 | Five seconds at the model's minimum resolution |
| Per character / per token (speech) | TTS models | $0.0145 /1K chars (Qwen TTS) · $1.5385 /1M tokens (MiniMax) | Audio output length |
| Per generation | 7 image models, 7 3D models, 3 audio models | $0.185 (Hunyuan 3D Motion) – $0.323 (Tripo 3D) | One render or one generated asset |
Seedance 2.5 is $0.70 per 5-second clip. That is the floor: five seconds at minimum resolution. A 1080p render, a longer clip or a premium mode costs proportionally more. If your product renders at high resolution, your real per-clip cost can be several times the headline number. Same for images: a "per image" price usually assumes a default size, not 4K.
MiniMax Hailuo and a few others bill in tokens computed from duration and resolution, not a fixed per-clip price. Their published "per token" rate looks tiny and incomparable to a per-clip rival until converted. We publish an estimated token rate and derive a per-5s-clip figure (cost_per_5s_est) so the two video families land in the same column.
Qwen TTS bills per character ($0.0145 /1K chars). MiniMax speech bills per token ($1.5385 /1M tokens). Characters and tokens are not convertible without a tokenizer assumption, and the ratio varies by language. A "cheaper per unit" claim across those two units is meaningless unless you fix a text sample and measure both.
Embedding models charge input only. A pricing table with two columns (input/output) makes them look artificially cheap or broken. In the normalized feed, price_per_1m_output is 0 for embeddings and the ranking uses input only.
Rather than pretend all prices are comparable, the feed exposes a common set of derived fields per model:
| Field | Meaning |
|---|---|
price_per_1m_input / price_per_1m_output | Text models, converted from per-1K or per-token rates |
cost_per_5s_est | Video models, converted from per-second or token rates at minimum resolution |
price_per_image | Image models, per generated image at default settings |
normalized_unit | Which of the above the row was normalized into (free for free models) |
price_usd / unit | The provider's own number and unit, unmodified — for audit |
Every ranking we publish (cheapest per task, cost per call, long-context cost) uses these derived fields, not the raw provider units. The calculator does the same arithmetic in the browser.
They measure different things. An image price covers one generation regardless of prompt length; a token price covers text volume. To compare, convert both to the cost of one typical job in your workload — that is what the normalized feed's derived fields are for.
It is the cheapest configuration of that video model: a five-second clip at its minimum resolution. Higher resolution, longer clips or premium modes cost proportionally more, so the headline unit price is a floor.
Yes. Some video models 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 so the number stays comparable.
Two different ways: per character (for example Qwen TTS at $0.0145 per 1K characters) or per token (MiniMax speech at $1.5385 per 1M tokens). Characters and tokens do not convert, so the two are not directly comparable.
111 models behind one OpenAI-compatible endpoint, 0% markup. The price feed is public — consume it or verify us with it.
Get an API keyUnits, prices and category counts reflect the live catalog on 2026-10-22. Raw numbers: /china-ai-leaderboard.json; free-tier limits: /transparency.html.