GPT Models from the OpenAI API
GPT models through the OpenAI API are in Preview: ready to use, and their settings may still change once a test on OpenAI's API confirms them.
The GPT models DocAI Fabric runs on Azure OpenAI can also come from your own OpenAI account: add GPT-6 Luna or GPT-6.1 Sol as your own model, with the profile DocAI Fabric already ships for it, and use it for any AI task. You need an API key from the OpenAI Platform and one tenant model on the Models page (Integrate > Models).
The profiles' settings were measured on Azure OpenAI, and OpenAI's own model pages give the same context window, output limit and APIs. The first Test on your account confirms them. The general rules for models and endpoints are in Models; what Test measures is in Testing and Measuring Models.
Before you start
- An OpenAI API key, created on the OpenAI Platform, on an organization with credit on it. API use is billed per token by OpenAI and is not part of a ChatGPT subscription.
- The
tenant.manage_modelspermission ("Manage AI models"). - A decision about where your documents may go. Requests go to OpenAI's API rather than to your own cloud region, with the document text and page images. If your tenant has data-residency commitments, confirm this route is allowed first.
The key is stored encrypted and cannot be read back. Keep your own copy.
Step 1: add the model
On the Models tab, click Add model.
| Field | What to enter |
|---|---|
| Model id | An id of your own, for example gpt-6-luna-own. It cannot be a system model's id, and gpt-6-luna, gpt-6-luna-no-reasoning, gpt-6.1-sol and gpt-6.2-sol are system models on DocAI Fabric's own service. OpenAI's spelling of the model goes into the endpoint's provider model name (Step 2). Immutable after creation. |
| Display name | Anything readable, for example GPT-6 Luna (own). |
| Profile | The dialog proposes gpt-6-luna, gpt-6.1-sol or gpt-6.2-sol from the id; keep it. It sets what the model refuses and needs, and its API: Chat Completions for Luna, Responses for the Sol models. |
| Context window (tokens) | 922000, the most one call may send to either model. |
| Supported tasks | Luna: split, classify, extract, redact, translate. Sol: assistant. |
| Modalities | Text and Image; for Sol on the assistant, Streaming too. |
| Max images per request | Leave blank: OpenAI's limit, 1,500, applies, and long documents reach the context window first. |
| Usage counted as and Estimated cost | input tokens and output tokens, at OpenAI's published rates for the model. |
Step 2: add the endpoint
| Field | Value |
|---|---|
| Name | Anything, for example openai. |
| Protocol | OpenAI-compatible. |
| Endpoint URL | https://api.openai.com |
| API path | /v1/chat/completions. The Sol profile calls the Responses API on the same endpoint, /v1/responses, by itself. |
| Provider model name | The model as OpenAI spells it, for example gpt-6-luna. |
| Authentication | Bearer token, with the API key as the value. |
da models list shows the endpoint's provider as openai, named from its host.
Step 3: test it, then put it to work
Use the Test link in the endpoint health column. It runs the operations the ticked tasks need and keeps the result on the model. A pass on text, structured output and images confirms the profile on your account; anything else, with the message Test shows, is what to send us: this page leaves Preview when a test on OpenAI's API confirms the settings.
Adding a model changes no project. Choose it as a task's default on the Defaults tab, or on one step of a project's workflow to try it there first.
What to expect in operation
- OpenAI's rate limits apply per organization, so a second endpoint with the same key adds no capacity.
- Usage and cost work as on any model: input and output tokens come back on every call and feed the Usage tab, budgets and cost estimates.