Models
GET /v1/models lists the models available to your organization right now.
beatrice-flash
The general-purpose model: polished Italian, coding, reasoning.
| Feature | Value |
|---|---|
| Id | beatrice-flash |
| Context | 262,144 tokens |
| Default output | 8,192 tokens plus the reasoning of the level |
Maximum output (max_tokens) | 49,152 tokens |
| Reasoning | yes, in reasoning_content |
| Tool calling | yes, parallel too |
- Reasoning: the model thinks before answering. Its reasoning comes in the
reasoning_contentfield (when streaming:delta.reasoning_content) and counts as output tokens. With amax_tokenstoo low the answer may stay empty because reasoning used all the room. - How much it reasons: set
reasoning_effort.noneanswers at once, without reasoning (minimalis the same);lowis the default;mediumreasons more;high(orxhigh) reasons thoroughly, is slower and uses more tokens. Reasoning stops at 2,048 tokens withlow, 8,192 withmediumand 16,384 withhigh; withoutmax_tokensthe room for the answer (8,192 tokens) comes on top of the reasoning. Any other value is a 400 error. On the Free plan the model always answers without reasoning, whatever level is asked. On a short question, measured on 2026-10-03: about 1.5 seconds withnone, 4 withmedium, 20 withhigh. - Cached context: when you continue a conversation the model has already seen, the repeated part is read from the cache. It is faster and does not count as input towards your limits (see Limits and plans).
beatrice-flash-extreme
The same model, without our automatic red-line check. It answers on its own, with its own refusal training.
| Feature | Value |
|---|---|
| Id | beatrice-flash-extreme |
| Model | the same as beatrice-flash |
| Limits | identical |
| Red lines | no check of ours |
- What changes: our five questions are not asked on the request or the answer, so you never get
finish_reason: "content_filter"from us. The 503safety_unavailableerror goes away too. - What does not change: the five red lines are still rules of use (Red lines) and on this model you answer for them. Violations are handled through reports and account suspension. Usage and limits are counted as on any other model.
- When to use it: when the check gets a legitimate job wrong, for example security research or legal and clinical texts, or when the last milliseconds of latency matter.
- Who can use it: verified organizations, that is students, professionals with a VAT number, companies, organizations and public bodies. You ask for the verification in the console, on the Plan page. On the paid plans it also needs the Extreme add-on; the Free plan does not include it. Without these,
GET /v1/modelsdoes not list it and a request gets403 model_not_allowed.
The public id stays stable even when the model behind it changes: updates and replacements are announced in advance.