A price list looks like a column you can sort. It is usually four different columns stacked on top of each other.
In the feed I pulled for this post, the Chat / LLM category contains 49 models across four billing units:
| Unit as printed | Models in that unit | What it actually means |
|---|---|---|
flat |
35 | Ambiguous. In practice almost all of these are per 1M tokens — see the defect note below |
per generation |
8 | A per-request price, not a per-token price |
per 1M tokens |
4 | Unambiguous |
Free |
2 | Rate-limited free tier |
Sort that column and you are comparing a per-token number against a per-request number. The sort is still a sort; it just is not a price ranking.
The four units, and when each is honest
Per 1M tokens. The only unit that scales with how much work you send. Right for chat, completion, reasoning, embeddings. If you are comparing chat models, this is the axis.
Per request / per generation. Right when the provider's cost is dominated by a fixed amount of work: one image, one video clip, one 3D asset. Not right for chat, because two chat requests can differ by 50x in token count.
Per image / per second / per megapixel. Media-specific. Per-megapixel is the only one that compares cleanly across models that output different resolutions.
Free. A price, not a unit. Requires knowing the rate limit to be comparable at all.
The defect, stated plainly
The feed this post is based on labels 35 Chat/LLM models as flat. Those models are billed per 1M tokens; the label is wrong. This is the same defect that has been recorded before, and it is still present in the current feed.
It matters because "flat" reads as per request to anyone scanning the table. A reader who sorts by price and compares a flat chat model against a per generation one is comparing two numbers that were never on the same axis.
Two more inconsistencies in the same feed, for completeness:
-
Embedding: 4 models
flat, 2 modelsper image— an embedding model priced per image is almost certainly a labelling error, since embeddings are billed by input tokens. -
Video: 11 models
per 1M tokens, 6per generation, 3per image— all three can be legitimate for video, which means video prices from different rows are not comparable without converting.
I am publishing the defect rather than the fix because the defect is the interesting part: a price table's failure mode is not being wrong, it is being sortable.
How to compare anyway
Group by unit before you sort. Never rank across units. Split the table by unit, rank inside each group, and only then decide whether the groups are substitutes for your use case.
Convert to the thing you actually buy. If you are building a chat feature, the unit you care about is cost per completed task, not cost per token. Those differ because of retries:
cost per completed task = (cost of all attempts) / (tasks that succeeded)
A model at $0.50/M tokens that needs three attempts costs the same as a model at $1.50/M that gets there in one. Per-token prices cannot tell you which is which; only your own pass rate can.
State the measurement window for anything time-dependent. Price is stable enough to publish with a date. Latency is not. In a three-timepoint probe on the same platform, one model measured:
| Time | TTFT median |
|---|---|
| 03:10 | 0.20s |
| 05:15 | 4.99s |
| 07:00 | 4.39s |
Same model, same endpoint, same prompt. That is 22x across four hours, while a sibling model from the same provider stayed inside 0.24–0.50s the whole time. Any latency number published without a measurement window is not comparable to anything.
A worked comparison from today's feed
Chat/LLM, price as printed (unit caveats as above):
| Model | price | Printed unit | |
|---|---|---|---|
| Cheapest | Hy4 Flash | 0.035 |
per generation ← should be per-1M-tokens |
| Hy-MT2 Lite | 0.043 | flat |
|
| Doubao Seed 2.0 Lite | 0.075 | flat |
|
| Qwen3.8 Max | 1.840 | flat |
|
| Kimi-K3.1-Pro | 2.000 | flat |
|
| Most expensive | Kimi-K3.1-Pro HighSpeed | 4.000 | flat |
Spread: 114.3x. Directionally that is the real story — a chat model choice spans two orders of magnitude. But do not read it as "model A is exactly 3.2x model B": six of these six rows carry a unit label I have just told you not to trust.
Category spreads, all units mixed, same caveat:
| Category | n priced | Spread (max ÷ min) |
|---|---|---|
| Chat / LLM | 47 | 114.3x |
| Audio | 11 | 106.2x |
| Image | 17 | 56.6x |
| Video | 21 | 28.5x |
| 3D | 9 | 3.0x |
| Embedding | 6 | 1.4x |
The shape is the useful finding. Chat and audio span two orders of magnitude; 3D and embedding span 1.4–3x. Price is a real lever for chat and almost irrelevant for 3D and embeddings — which is the opposite of how much time people spend comparing them.
Summary
- Group by unit before ranking. Four units live inside the Chat category alone.
-
flatis not a unit. On this feed it means per-1M-tokens for chat and is unresolved elsewhere. - Per-token price does not predict cost per completed task. Retries are in the second number and not the first.
- Never publish or trust a latency figure without its measurement window. 22x in four hours on one route.
- Spread is a shape, not a ranking. 114x in chat tells you where the decision is worth making; it does not tell you which model is 3x another.
Prices: public feed, generated_at 2026-11-19 19:00 UTC, 116 models, 0% platform markup. Latency: 3 samples per model per timepoint, single region, streaming TTFT, median. Feed: https://aiapi-pro.com/china-ai-leaderboard.json.
Top comments (0)