Models
Two models, one cheap and one careful. The numbers below are measured on this service, not copied from a vendor page.
What is available
| MODEL | TOKENS PER SHORT REPLY | REPLIES PER AD |
|---|---|---|
| — | Loading | — |
Both are served through a third-party provider and both take the same request shape. Swapping between them is one string in your code.
Which one to pick
The gap between them is about five times the cost per reply. That is the whole decision.
Use the cheap one for
- Classification, tagging, yes-or-no questions.
- Rewriting, summarising, translating short text.
- Anything you are going to run in a loop over many rows.
- Prototyping, while you are still changing the prompt every two minutes.
Use the careful one for
- Multi-step reasoning where a wrong middle step ruins the answer.
- Code that has to run, not code that has to look right.
- Long instructions with several constraints at once.
- The final pass, after the cheap one got you most of the way.
A practical pattern: draft with the cheap one, and only send the cases it got wrong to the careful one. On a free balance that stretches a lot further than sending everything to the expensive model.
They think before answering
Both models write out a reasoning pass before the visible answer. You do not see it, but it is counted, and it is most of what you pay for on a short question.
Loading
If a reply comes back empty, this is almost always why: the cap was spent on thinking before the answer began. Leaving the output limit unset is the safest default.
How far one ad goes
Loading
| WHAT YOU ARE DOING | ROUGH COST |
|---|---|
| One short question | 150 to 750, depending on the model. |
| A long document to summarise | The document itself counts too. A 4,000-word page is roughly 6,000 before the answer starts. |
| A coding agent editing a file | Several thousand per step, because it resends the file and its instructions every time. |
| A back-and-forth chat | Grows each turn — the whole history is resent every message. |
What is not here
No image input, no image generation, no audio, no embeddings, no function calling guarantees. This is a free beta running two chat models. If your code needs one of those, this is not the endpoint for it.
The list can change without notice. Read it from the models endpoint at runtime rather than hard-coding names.