Custom endpoint routing
Requests must pass through an internal gateway or OpenAI-compatible proxy.
AI Tools · Function
createOpenAI
createOpenAI creates a customized OpenAI provider for the AI SDK. Configure credentials, endpoint, headers, or transport, then use the provider to create OpenAI models.
How it works
createOpenAI builds an OpenAI provider instance from optional settings such as credentials, endpoint, headers, and fetch implementation. The resulting provider is callable with a model ID and also exposes provider-specific model factories such as chat when needed.
Syntax
createOpenAI(options?: { apiKey?: string; baseURL?: string; headers?: Record<string, string>; fetch?: typeof fetch }): OpenAIProviderParameters
| Name | Type | Description |
|---|---|---|
options | object | Optional provider configuration, including baseURL, apiKey, headers, and fetch. |
Returns
A configured OpenAI provider instance that can create models from model IDs.
Key points
- Import
createOpenAIfrom@ai-sdk/openaiwhen the default provider instance is not sufficient. - Set
baseURLto route requests through a proxy or another OpenAI-compatible endpoint. - Without an explicit base URL, resolution falls back first to
OPENAI_BASE_URLand then to the documented OpenAI endpoint. - The configured API key is transmitted through the
Authorizationheader. - If no key is passed in settings, the provider reads
OPENAI_API_KEY. - Factory settings can include a custom
fetchimplementation as well as a custom base URL.
Common mistakes
- The default factory targets OpenAI's Responses API; use the
chatfactory when the endpoint only supports Chat Completions. - A custom endpoint that lacks Responses API support requires
openai.chat('model-id')rather than the default model factory.
Example: Create a configured OpenAI provider
import { createOpenAI } from '@ai-sdk/openai';
const openai = createOpenAI({
apiKey: process.env.OPENAI_API_KEY,
});
const model = openai('gpt-5');Creates a configured OpenAI provider and then obtains a model instance by model ID.
When to use it
Per-instance authentication
Different deployments need separate credentials or request headers.
Custom transport
A runtime needs instrumentation, retries, or a WebSocket-backed request implementation.
Interview questions
QWhat does createOpenAI provide?
It returns a configured OpenAI provider instance. Configuration can override the API key, base URL, headers, and fetch implementation, while unspecified credentials and the standard endpoint can come from environment variables or defaults.
Follow-up: How would you target a proxy that implements only Chat Completions? Use baseURL for the proxy and create the model through the provider's chat factory.
QHow is a provider created and then used?
The factory is imported from @ai-sdk/openai and accepts an optional settings object. The returned provider can be called with a model ID, for example openai('model-id').
Follow-up: How can the provider use a custom transport? Supply a custom fetch function in the factory settings.
QWhat are the credential and endpoint defaults?
The API key setting is sent in the Authorization header. If omitted, the provider uses OPENAI_API_KEY; the base URL similarly defaults through OPENAI_BASE_URL and then the standard OpenAI v1 endpoint.
Follow-up: What should be checked when a proxy behaves unexpectedly? Verify both the configured base URL and whether it supports the API selected by the model factory.
Related
openaiopenai.chat@ai-sdk/openaistreamTextgenerateText