Answer Models
An answer model is a model Ask writes answers with, on your own provider key, with an optional monthly token budget.
- Set up several and pick one per request — a cheap model for one audience, a stronger one for another.
- Mark one default and Ask uses it whenever a request names neither a model nor a judge.
- Every answer reports the tokens it cost, so you can price answers.
Add an answer model
- Go to Models → Answer Models and click Add Answer Model.
- Pick a provider — Anthropic, OpenAI, Gemini or Microsoft Foundry — and a model. Models released since the list was last updated work too: type the name. For Microsoft Foundry, type your deployment’s name exactly as Azure spells it: any model you have deployed works (Grok, DeepSeek, Llama, Mistral, Phi, MAI and more).
- Choose a saved credential or paste a new API key.
- Click Add. One short call checks that the model answers on that key before it is saved.
One answer model per provider model per account.
Choose it on a request
Pass the name, {provider}-{model}, as model:
{ "question": "which jackets are waterproof?", "model": "anthropic-claude-haiku-4-5" }A Foundry model is microsoft-{deployment name}, e.g. microsoft-Grok-4-Fast. Any casing matches; the deployment name reaches Foundry as you saved it.
| Request has | Ask answers with |
|---|---|
model | That answer model. Not one of yours → an error listing the ones you have. |
judge, no model | That judge’s model and key. |
| Neither | Your default answer model, else your only judge. |
An account with no answer models answers exactly as before.
What it costs
- Each response carries
usage:{ "input_tokens", "output_tokens" }, as your provider counted them, every attempt included. The streameddoneevent carries the same. - The model’s row shows this month’s tokens (UTC).
| Provider | Input counts | Output counts |
|---|---|---|
| Anthropic | input_tokens + cache reads + cache writes | output_tokens |
| OpenAI | prompt_tokens | completion_tokens (reasoning included) |
| Gemini | promptTokenCount | candidatesTokenCount + thoughtsTokenCount |
| Microsoft Foundry | prompt_tokens | completion_tokens |
Monthly budget
- Monthly budget… on a model’s ⋯ menu caps the tokens (input plus output) it may spend in a calendar month (UTC). Empty means no limit.
- At the cap, Ask still searches and returns the records, with
answered: falseandreason: "budget". No model is called. - The count resets on the 1st. Raise the budget to answer again sooner.
Streaming
Every provider’s answer models stream word by word on /stream, Microsoft Foundry included.
Delete
- Nothing is attached to an answer model, so it deletes at once. A request still naming it gets the error listing the models left.
- A credential an answer model uses cannot be deleted until the model is.