Models

Model ids

Use the author/name form exactly as listed. A deprecated model stays in this table until its retirement date, with the date and its replacement; see Deprecation and retirement below.

Deprecation and retirement

A model is deprecated with at least 30 days’ notice before it retires. Until the retirement date it keeps working as before, and every response to it carries these headers:

Deprecation: @1788220800                   # when the deprecation was announced (RFC 9745: @ and Unix seconds)
Sunset: Thu, 03 Dec 2026 00:00:00 GMT      # when the model retires (RFC 8594: an HTTP date)
Link: <https://…/notice>; rel="deprecation"  # the published notice, when there is one

Owners, billing admins and admins are told when a model is deprecated, 30 and 7 days before it retires, and when it retires; so are the people who created the keys that called it in the last 30 days, and the admins of those keys’ projects. The console overview lists the affected keys and the replacement.

On the retirement date, a model with a redirect is served by its replacement: the response names the model that answered and x-tf-model-redirect the one you asked for. A key whose allowed models include the retired model may call its replacement through the redirect. Without a redirect, requests return 410 model_retired.

Model aliases

Call @name instead of a model ID, and change which model answers without redeploying. Project admins (and owners and admins) define a project’s aliases under Settings → Model aliases; owners and admins define organization aliases, which every project can use, under Settings → Models. A project alias overrides an organization alias with the same name. Names use lower-case letters, digits, dots, dashes and underscores.

curl https://gridport.ai/v1/chat/completions -H "Authorization: Bearer $GRIDPORT_API_KEY" \
  -d '{"model": "@cs-default", "messages": [{"role": "user", "content": "Hello"}]}'
# x-tf-model-alias in the response names the alias; "model" names the model that answered

An alias has up to five targets, each a model (optionally pinned as model@version) with a weight from 1 to 100. Each request goes to one target in proportion to the weights, and the same request ID always goes to the same target, so a retry lands where the first attempt did. Moving weight from one target to another rolls a new model out gradually.

A key’s allowed models may list the alias, which allows every one of its targets, or the target models themselves. An alias that does not exist returns 404 and says where to define it. GET /v1/models lists the aliases a key can use.

The organization’s model list

Owners and admins can name the models anyone in the organization may call, under Settings → Models. With a list in place, every key’s allowed models must come from it, and a key may call the models on both lists; a key without a list of its own may call every model on the organization’s. An empty organization list allows every model. A refusal says whether the organization or the key does not allow the model, and GET /v1/models lists exactly the models the key can call. Members, roles and project access