Almost every conversation about AI models is about chat models. Of the 354 we track, 227 are chat. The other 127 do different work — and they are not even priced in the same unit.
How many of each
| Kind | Models | Median input |
|---|---|---|
| Chat | 227 | $0.50 |
| Transcription | 53 | $0.00 |
| Embedding | 41 | $0.10 |
| Rerank | 13 | $2.00 |
| Speech | 11 | $0.03 |
| Video | 9 | - |
Rerank costs more than chat
The number that stands out is rerank: a median of $2.00 per million tokens, four times the $0.50 of chat models. Reranking just reorders search results — far simpler work than conversation — yet it costs more.
The likely reason is volume. Reranking has to take in every candidate document, so each call carries many tokens, and it runs on every search. Even at the same rate the bill grows. That also leaves vendors room to charge more here.
Transcription has a median of zero
The median input price across 53 transcription models is $0.00. That does not mean they are free. This kind is usually billed by the second or minute rather than by token, so the token column is simply empty. For the same reason the nine video models have no median at all.
This is also a limit of our table: values measured in different units cannot share a column. So each model page states what the rate counts alongside the number itself. When choosing a speech or video model, read that unit first.
Embeddings are cheap but used in bulk
The 41 embedding models have a median of $0.10, a fifth of chat. But embeddings run over an entire corpus at indexing time, and again whenever documents change. The rate is low while the volume is high, so totals overshoot expectations easily. Here the useful question is not "cost per call" but "cost to embed everything I have."