Supported model providers¶

Every pydantic-ai model call in this provider, from an operator, a decorator or the hook, reaches its model through one Airflow connection, and the provider: prefix on the model name picks the vendor. Switching vendors is a connection change, not a Dag change. Find your vendor below, install the extra, create the connection type shown, and set the model name with that prefix.

For the exhaustive, generated-and-hand-written list of every service this provider reaches (connections and toolsets alike), see Supported services.

Vendor	Model prefix	Install	Connection type	Credentials
OpenAI	`openai:`	`[openai]`	`pydanticai`	API key in Password
Anthropic	`anthropic:`	`[anthropic]`	`pydanticai`	API key in Password
Google Gemini API	`google:`	`[google]`	`pydanticai`	API key in Password, or `GOOGLE_API_KEY` in the environment
Google Vertex AI	`google-cloud:`	`[google]`	`pydanticai_vertex` (Pydantic AI (Google Vertex AI) connection)	Service account or Application Default Credentials
AWS Bedrock	`bedrock:`	`[bedrock]`	`pydanticai_bedrock` (Pydantic AI (AWS Bedrock) connection), or `pydanticai` with AWS credentials in the environment	IAM keys, profile or role
Azure OpenAI	`azure:`	`[openai]`	`pydanticai_azure` (Pydantic AI (Azure OpenAI) connection)	API key in Password, resource endpoint in Host
Groq	`groq:`	`pydantic-ai-slim[groq]`	`pydanticai`	API key in Password
Mistral AI	`mistral:`	`pydantic-ai-slim[mistral]`	`pydanticai`	API key in Password
DeepSeek	`deepseek:`	`[openai]`	`pydanticai`	API key in Password
Ollama, vLLM and other OpenAI-compatible servers	`openai:`	`[openai]`	`pydanticai` with the server URL in Host (Self-hosted models)	Usually none
Snowflake Cortex	`snowflake:`	`pydantic-ai-slim[snowflake]` for `pydanticai`, or `apache-airflow-providers-snowflake[common.ai]` for `pydanticai_snowflake`	`pydanticai`, or `pydanticai_snowflake` (Pydantic AI (Snowflake Cortex) connection)	`SNOWFLAKE_ACCOUNT` and `SNOWFLAKE_TOKEN` in the worker environment; leave Password empty. The `pydanticai_snowflake` connection type instead sources OAuth, PAT or key-pair credentials from an existing Snowflake connection.
TypeSafe Jev (decision model, does not write text)	`typesafe:`	`[typesafe]`	`pydanticai` (Decision models)	API key in Password
Decision models behind the System One API: Ollama, Strands Decider, Kev and others (do not write text)	`system-one:`	None (`pydantic-ai-slim` 2.53.0+)	`pydanticai` with the server URL in Host (Decision models)	API key in Password, if the server takes one

[name] in the Install column is an extra of this provider, installed as pip install "apache-airflow-providers-common-ai[name]". The Groq and Mistral entries name the matching pydantic-ai-slim extra instead, because this provider does not ship one for them. The Snowflake entry lists one install per connection type: pydanticai uses the pydantic-ai-slim extra, while the pydanticai_snowflake type ships with the Snowflake provider and is installed through its [common.ai] extra.

Any other vendor that pydantic-ai supports (Cohere, OpenRouter, Hugging Face and more) works the same way: install the pydantic-ai-slim extra named on that vendor’s pydantic-ai page, create a pydanticai connection, and use the prefix from that page.

Hosted OpenAI-compatible services¶

A hosted service that exposes the OpenAI-compatible API is set up like a self-hosted server: a pydanticai connection with the service’s base URL in Host, its API key in Password, and the openai: prefix on the model name. For example, TrustedRouter is an OpenAI-compatible router:

{
    "conn_type": "pydanticai",
    "host": "https://api.trustedrouter.com/v1",
    "password": "<api-key>",
    "extra": "{\"model\": \"openai:trustedrouter/auto\"}"
}

Where the model name goes¶

Set the model on the connection’s Model field in provider:model form, for example anthropic:claude-sonnet-5. An operator’s model_id overrides it for that task, so one connection can serve several models from the same vendor. The generic pydanticai connection has no vendor of its own, so a bare model name without a prefix is rejected; the Azure, Bedrock and Vertex connection types each supply their own prefix and accept a bare name. Pydantic AI connection has every field and the full resolution order.

Reliability across vendors¶

A connection can name other connections to fail over to when its vendor is unavailable, so one task can span OpenAI and Anthropic without a code change. See Provider fallback.

Using a model outside an operator¶

Using the hook directly: PydanticAIHook returns the pydantic-ai Agent or Model behind a connection for use in a plain @task. LangChain models: LangChainHook does the same for LangChain chat and embedding models, which have their own langchain connection type.

Supported model providers¶

Hosted OpenAI-compatible services¶

Where the model name goes¶

Reliability across vendors¶

Using a model outside an operator¶

Pages in this section¶