openai

openai/openai-python

The official Python library for the OpenAI API

View on GitHub Website / Demo Back to listing

32kstars 6.6kforks 100+contributors Python

README.md

# OpenAI Python API library

[![PyPI version](https://img.shields.io/pypi/v/openai.svg?label=pypi%20(stable))](https://pypi.org/project/openai/)

The OpenAI Python library provides convenient access to the OpenAI REST API from any Python 3.10+
application. The library includes type definitions for all request params and response fields,
and offers both synchronous and asynchronous clients powered by [HTTPX2](https://httpx2.pydantic.dev/).

It is generated from our [OpenAPI specification](https://github.com/openai/openai-openapi).

## Documentation

The REST API documentation can be found on [platform.openai.com](https://platform.openai.com/docs/api-reference). The full API of this library can be found in [api.md](api.md).

## Installation

```sh
# install from PyPI
pip install openai
```

## Usage

The full API of this library can be found in [api.md](api.md).

The primary API for interacting with OpenAI models is the [Responses API](https://developers.openai.com/api/reference/resources/responses). You can generate text from the model with the code below.

```python
import os
from openai import OpenAI

client = OpenAI(
    # This is the default and can be omitted
    api_key=os.environ.get("OPENAI_API_KEY"),
)

response = client.responses.create(
    model="gpt-5.5",
    instructions="You are a coding assistant that talks like a pirate.",
    input="How do I check if a Python object is an instance of a class?",
)

print(response.output_text)
```

The previous standard (supported indefinitely) for generating text is the [Chat Completions API](https://platform.openai.com/docs/api-reference/chat). You can use that API to generate text from the model with the code below.

```python
from openai import OpenAI

client = OpenAI()

completion = client.chat.completions.create(
    model="gpt-5.5",
    messages=[
        {"role": "developer", "content": "Talk like a pirate."},
        {
            "role": "user",
            "content": "How do I check if a Python object is an instance of a class?",
        },
    ],
)

print(completion.choices[0].message.content)
```

While you can provide an `api_key` keyword argument,
we recommend using [python-dotenv](https://pypi.org/project/python-dotenv/)
to add `OPENAI_API_KEY="My API Key"` to your `.env` file
so that your API key is not stored in source control.
[Get an API key here](https://platform.openai.com/settings/organization/api-keys).

### Workload Identity Authentication

For secure, automated environments like cloud-managed Kubernetes, Azure, and Google Cloud Platform, you can use workload identity authentication with short-lived tokens from cloud identity providers instead of long-lived API keys.

#### Kubernetes (service account tokens)

```python
from openai import OpenAI
from openai.auth import k8s_service_account_token_provider

client = OpenAI(
    workload_identity={
        "identity_provider_id": "idp-123",
        "service_account_id": "sa-456",
        "provider": k8s_service_account_token_provider("/var/run/secrets/kubernetes.io/serviceaccount/token"),
    },
)

response = client.chat.completions.create(
    model="gpt-5.5",
    messages=[{"role": "user", "content": "Hello!"}],
)
```

#### Azure (managed identity)

```python
from openai import OpenAI
from openai.auth import azure_managed_identity_token_provider

client = OpenAI(
    workload_identity={
        "identity_provider_id": "idp-123",
        "service_account_id": "sa-456",
        "provider": azure_managed_identity_token_provider(
            resource="https://management.azure.com/",
        ),
    },
)
```

#### Google Cloud Platform (compute engine metadata)

```python
from openai import OpenAI
from openai.auth import gcp_id_token_provider

client = OpenAI(
    workload_identity={
        "identity_provider_id": "idp-123",
        "service_account_id": "sa-456",
        "provider": gcp_id_token_provider(audience="https://api.openai.com/v1"),
    },
)
```

#### Custom subject token provider

```python
from openai import OpenAI

def get_custom_token() -> str:
    return "your-jwt-token"

client = OpenAI(
    workload_identity={
        "identity_provider_id": "idp-123",
        "service_account_id": "sa-456",
        "provider": {
            "token_type": "jwt",
            "get_token": get_custom_token,
        },
    }
)
```

You can also customize the token refresh buffer (default is 1200 seconds (20 minutes) before expiration):

```python
from openai import OpenAI
from openai.auth import k8s_service_account_token_provider

client = OpenAI(
    workload_identity={
        "identity_provider_id": "idp-123",
        "service_account_id": "sa-456",
        "provider": k8s_service_account_token_provider("/var/token"),
        "refresh_buffer_seconds": 120.0,
    }
)
```

#### X.509 workload identity (mutual TLS)

For X.509 workload identity federation, configure the client certificate and
server trust on an HTTPX2 client, then pass only the identity-provider and
service-account IDs to the SDK:

```python
import os
import ssl

from openai import OpenAI, DefaultHttpx2Client
from openai.auth import x509_workload_identity

tls_context = ssl.create_default_context(
    cafile=os.getenv("OPENAI_MTLS_CA_BUNDLE"),
)
tls_context.load_cert_chain(
    certfile=os.environ["OPENAI_MTLS_CERTIFICATE_CHAIN"],
    keyfile=os.environ["OPENAI_MTLS_PRIVATE_KEY"],
    password=os.getenv("OPENAI_MTLS_PRIVATE_KEY_PASSWORD"),
)

client = OpenAI(
    workload_identity=x509_workload_identity(
        identity_provider_id=os.environ["OPENAI_IDENTITY_PROVIDER_ID"],
        service_account_id=os.environ["OPENAI_SERVICE_ACCOUNT_ID"],
        # refresh_buffer_seconds=120.0,
    ),
    http_client=DefaultHttpx2Client(
        verify=tls_context,
        follow_redirects=False,
    ),
)
```

X.509 mode defaults to `https://mtls.api.openai.com/v1` when neither `base_url`
nor `OPENAI_BASE_URL` is set. The same configured HTTP client presents its
certificate to the fixed mTLS token-exchange endpoint and to the API. Tokens
are exchanged lazily, cached, and refreshed automatically. Certificate files,
private keys, passwords, server trust, proxies, and rotation remain application
and transport concerns.

X.509 API requests require HTTPS and must stay on the configured API origin.
The effective HTTP Host authority must match that origin.
Provider API-key and proxy-only headers cannot be sent to the API alongside
X.509 authentication.
Token exchanges do not inherit API request hooks, authentication, or cookies.
Identity settings are captured when the client is constructed; create a new
client to change the identity. Azure clients do not support X.509 workload
identity.

For asynchronous requests, use `AsyncOpenAI` with
`DefaultAsyncHttpx2Client`. See the complete [sync rollout-toggle
example](examples/x509_workload_identity.py) and [async rollout-toggle
example](examples/x509_workload_identity_async.py), which select API-key or
X.509 authentication with the application-owned `OPENAI_AUTH_MODE`
environment variable. X.509 workload identity currently supports HTTP APIs;
Realtime and WebSockets are not included.

### Vision

With an image URL:

```python
prompt = "What is in this image?"
img_url = "https://api.nga.gov/iiif/a2e6da57-3cd1-4235-b20e-95dcaefed6c8/full/!800,800/0/default.jpg"

response = client.responses.create(
    model="gpt-5.5",
    input=[
        {
            "role": "user",
            "content": [
                {"type": "input_text", "text": prompt},
                {"type": "input_image", "image_url": f"{img_url}"},
            ],
        }
    ],
)
```

With the image as a base64 encoded string:

```python
import base64
from openai import OpenAI

client = OpenAI()

prompt = "What is in this image?"
with open("path/to/image.png", "rb") as image_file:
    b64_image = base64.b64encode(image_file.read()).decode("utf-8")

## Async usage

Simply import `AsyncOpenAI` instead of `OpenAI` and use `await` with each API call:

```python
import os
import asyncio
from openai import AsyncOpenAI

client = AsyncOpenAI(
    # This is the default and can be omitted
    api_key=os.environ.get("OPENAI_API_KEY"),
)

async def main() -> None:
    response = await client.responses.create(
        model="gpt-5.5", input="Explain disestablishmentarianism to a smart five year old."
    )
    print(response.output_text)

asyncio.run(main())
```

Functionality between the synchronous and asynchronous clients is otherwise identical.

### With aiohttp

By default, the async client uses HTTPX2. For improved concurrency performance, you may also use `aiohttp` as the HTTPX2 transport.

You can enable this by installing `aiohttp`:

```sh
# install from PyPI
pip install openai[aiohttp]
```

Then you can enable it by instantiating the client with `http_client=DefaultAioHttpClient()`:

```python
import os
import asyncio
from openai import DefaultAioHttpClient
from openai import AsyncOpenAI

async def main() -> None:
    async with AsyncOpenAI(
        api_key=os.environ.get("OPENAI_API_KEY"),  # This is the default and can be omitted
        http_client=DefaultAioHttpClient(),
    ) as client:
        chat_completion = await client.chat.completions.create(
            messages=[
                {
                    "role": "user",
                    "content": "Say this is a test",
                }
            ],
            model="gpt-5.5",
        )

asyncio.run(main())
```

### HTTPX2 migration

HTTPX2 is the default HTTP client. If you configure a custom HTTP client, transport, timeout, authentication handler, event hook, or request mock, see the [HTTPX2 migration guide](httpx2.md).

## Streaming responses

We provide support for streaming responses using Server-Sent Events (SSE).

```python
from openai import OpenAI

client = OpenAI()

stream = client.responses.create(
    model="gpt-5.5",
    input="Write a one-sentence bedtime story about a unicorn.",
    stream=True,
)

for event in stream:
    print(event)
```

The async client uses the exact same interface.

```python
import asyncio
from openai import AsyncOpenAI

client = AsyncOpenAI()

async def main():
    stream = await client.responses.create(
        model="gpt-5.5",
        input="Write a one-sentence bedtime story about a unicorn.",
        stream=True,
    )

async for event in stream:
        print(event)

asyncio.run(main())
```

## Realtime API

The Realtime API enables you to build low-latency, multi-modal conversational experiences. It currently supports text and audio as both input and output, as well as [function calling](https://platform.openai.com/docs/guides/function-calling) through a WebSocket connection.

Under the hood the SDK uses the [`websockets`](https://websockets.readthedocs.io/en/stable/) library to manage connections.

The retained `client.beta.realtime` interface does not exchange workload-identity tokens or sign Bedrock requests. A client using either method must supply its own `Authorization` settings through `default_headers` or `connect(extra_headers=...)`. Otherwise, the SDK rejects the WebSocket connection locally.

Connection headers override default headers case-insensitively, and `omit` removes a header. These settings preserve the caller's header choices without enabling provider authentication. X.509 workload identity supports HTTP only. Beta WebSockets always reject X.509 clients, even with explicit headers.

The Realtime API works through a combination of client-sent events and server-sent events. Clients can send events to do things like update session configuration or send text and audio inputs. Server events confirm when audio responses have completed, or when a text response from the model has been received. A full event reference can be found [here](https://platform.openai.com/docs/api-reference/realtime-client-events) and a guide can be found [here](https://platform.openai.com/docs/guides/realtime).

Basic text based example:

```py
import asyncio
from openai import AsyncOpenAI, OpenAIError

async def main():
    client = AsyncOpenAI()

async with client.realtime.connect(model="gpt-realtime-2") as connection:
        await connection.session.update(session={"type": "realtime", "output_modalities": ["text"]})

await connection.conversation.item.create(
            item={
                "type": "message",
                "role": "user",
                "content": [{"type": "input_text", "text": "Say hello!"}],
            }
        )
        await connection.response.create()

async for event in connection:
            if event.type == "response.output_text.delta":
                print(event.delta, flush=True, end="")

elif event.type == "response.output_text.done":
                print()

elif event.type == "response.done":
                if event.response.status != "completed":
                    raise OpenAIError("The Realtime response did not complete")
                break
            elif event.type == "error":
                raise OpenAIError("Realtime returned an error; inspect the event in your application")
        else:
            raise OpenAIError("Realtime disconnected before response.done")

asyncio.run(main())
```

However the real magic of the Realtime API is handling audio inputs / outputs, see this example [TUI script](https://github.com/openai/openai-python/blob/main/examples/realtime/push_to_talk_app.py) for a fully fledged example.

### Realtime error handling

Whenever an error occurs, the Realtime API will send an [`error` event](https://platform.openai.com/docs/guides/realtime-model-capabilities#error-handling) and the connection will stay open and remain usable. This means you need to handle it yourself, as _no errors are raised directly_ by the SDK when an `error` event comes in.

```py
client = AsyncOpenAI()

async with client.realtime.connect(model="gpt-realtime-2") as connection:
    ...
    async for event in connection:
        if event.type == "error":
            print(event.error.type)
            print(event.error.code)
            print(event.error.event_id)
            print(event.error.message)
```

Use the [Translation session helper](docs/realtime-translation.md) to send audio, finish input, and receive final events.

## Using types

Nested request parameters are [TypedDicts](https://docs.python.org/3/library/typing.html#typing.TypedDict). Responses are [Pydantic models](https://docs.pydantic.dev) which also provide helper methods for things like:

- Serializing back into JSON, `model.to_json()`
- Converting to a dictionary, `model.to_dict()`

Typed requests and responses provide autocomplete and documentation within your editor. If you would like to see type errors in VS Code to help catch bugs earlier, set `python.analysis.typeCheckingMode` to `basic`.

## Pagination

List methods in the OpenAI API are paginated.

This library provides auto-paginating iterators with each list response, so you do not have to request successive pages manually:

```python
from openai import OpenAI

client = OpenAI()

all_jobs = []
# Automatically fetches more pages as needed.
for job in client.fine_tuning.jobs.list(
    limit=20,
):
    # Do something with job here
    all_jobs.append(job)
print(all_jobs)
```

Or, asynchronously:

```python
import asyncio
from openai import AsyncOpenAI

client = AsyncOpenAI()

async def main() -> None:
    all_jobs = []
    # Iterate through items across all pages, issuing requests as needed.
    async for job in client.fine_tuning.jobs.list(
        limit=20,
    ):
        all_jobs.append(job)
    print(all_jobs)

asyncio.run(main())
```

Alternatively, you can use the `.has_next_page()`, `.next_page_info()`, or `.get_next_page()` methods for more granular control working with pages:

```python
first_page = await client.fine_tuning.jobs.list(
    limit=20,
)
if first_page.has_next_page():
    print(f"will fetch next page using these details: {first_page.next_page_info()}")
    next_page = await first_page.get_next_page()
    print(f"number of items we just fetched: {len(next_page.data)}")

# Remove `await` for non-async usage.
```

Or just work directly with the returned data:

```python
first_page = await client.fine_tuning.jobs.list(
    limit=20,
)

print(f"next page cursor: {first_page.after}")  # => "next page cursor: ..."
for job in first_page.data:
    print(job.id)

# Remove `await` for non-async usage.
```

## Nested params

Nested parameters are dictionaries, typed using `TypedDict`, for example:

```python
from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    input=[
        {
            "role": "user",
            "content": "How much ?",
        }
    ],
    model="gpt-5.5",
    text={"format": {"type": "json_object"}},
)
```

## File uploads

Request parameters that correspond to file uploads can be passed as `bytes`, a file-like object, a [`PathLike`](https://docs.python.org/3/library/os.html#os.PathLike) instance, or a tuple of `(filename, contents, media type)`.

```python
from pathlib import Path
from openai import OpenAI

client = OpenAI()

client.files.create(
    file=Path("input.jsonl"),
    purpose="fine-tune",
)
```

When uploading an in-memory file-like object such as `io.BytesIO`, include a filename when the API needs the file extension to determine its format. Passing a tuple is the most explicit option:

```python
import io

audio = io.BytesIO(audio_bytes)
transcription = client.audio.transcriptions.create(
    model="gpt-4o-transcribe",
    file=("audio.wav", audio, "audio/wav"),
)
```

You can also set a `.name` attribute such as `audio.name = "audio.wav"` on a mutable file-like object before passing it directly. Without a filename, multipart transports may use a generic name such as `upload`, which does not provide an audio extension for format detection.

The async client uses the exact same interface. If you pass a [`PathLike`](https://docs.python.org/3/library/os.html#os.PathLike) instance, the file contents will be read asynchronously automatically.

## Webhook Verification

Verifying webhook signatures is _optional but encouraged_.

For more information about webhooks, see [the API docs](https://platform.openai.com/docs/guides/webhooks).

### Parsing webhook payloads

For most use cases, you will likely want to verify the webhook and parse the payload at the same time. To achieve this, we provide the method `client.webhooks.unwrap()`, which parses a webhook request and verifies that it was sent by OpenAI. This method will raise an error if the signature is invalid.

Note that the `body` parameter must be the raw JSON string sent from the server (do not parse it first). The `.unwrap()` method will parse this JSON for you into an event object after verifying the webhook was sent from OpenAI.

```python
from openai import OpenAI
from flask import Flask, request

app = Flask(__name__)
client = OpenAI()  # OPENAI_WEBHOOK_SECRET environment variable is used by default

@app.route("/webhook", methods=["POST"])
def webhook():
    request_body = request.get_data(as_text=True)

try:
        event = client.webhooks.unwrap(request_body, request.headers)

if event.type == "response.completed":
            print("Response completed:", event.data)
        elif event.type == "response.failed":
            print("Response failed:", event.data)
        else:
            print("Unhandled event type:", event.type)

return "ok"
    except Exception as e:
        print("Invalid signature:", e)
        return "Invalid signature", 400

if __name__ == "__main__":
    app.run(port=8000)
```

### Verifying webhook payloads directly

In some cases, you may want to verify the webhook separately from parsing the payload. If you prefer to handle these steps separately, we provide the method `client.webhooks.verify_signature()` to _only verify_ the signature of a webhook request. Like `.unwrap()`, this method will raise an error if the signature is invalid.

Note that the `body` parameter must be the raw JSON string sent from the server (do not parse it first). You will then need to parse the body after verifying the signature.

```python
import json
from openai import OpenAI
from flask import Flask, request

app = Flask(__name__)
client = OpenAI()  # OPENAI_WEBHOOK_SECRET environment variable is used by default

@app.route("/webhook", methods=["POST"])
def webhook():
    request_body = request.get_data(as_text=True)

try:
        client.webhooks.verify_signature(request_body, request.headers)

# Parse the body after verification
        event = json.loads(request_body)
        print("Verified event:", event)

return "ok"
    except Exception as e:
        print("Invalid signature:", e)
        return "Invalid signature", 400

if __name__ == "__main__":
    app.run(port=8000)
```

## Handling errors

When the library is unable to connect to the API (for example, due to network connection problems or a timeout), a subclass of `openai.APIConnectionError` is raised.

When the API returns a non-success status code (that is, 4xx or 5xx
response), a subclass of `openai.APIStatusError` is raised, containing `status_code` and `response` properties.

All errors inherit from `openai.APIError`.

When consuming a `Stream` or `AsyncStream`, read timeouts raise `APITimeoutError`
and other HTTPX request failures raise `APIConnectionError`. Catch these SDK
exceptions instead of raw HTTPX exceptions; the original exception is available
as `__cause__`. Stream consumption is not automatically retried, because replaying
a request could duplicate output already delivered to your application.
The Assistants event-handler helpers and raw `with_streaming_response` iterators
retain their existing exception behavior.

```python
import openai
from openai import OpenAI

client = OpenAI()

try:
    client.fine_tuning.jobs.create(
        model="gpt-4o",
        training_file="file-abc123",
    )
except openai.APIConnectionError as e:
    print("The server could not be reached")
    print(e.__cause__)  # an underlying Exception, likely raised within HTTPX2.
except openai.RateLimitError as e:
    print("A 429 status code was received; we should back off a bit.")
except openai.APIStatusError as e:
    print("Another non-200-range status code was received")
    print(e.status_code)
    print(e.response)
```

Error codes are as follows:

| Status Code | Error Type                 |
| ----------- | -------------------------- |
| 400         | `BadRequestError`          |
| 401         | `AuthenticationError`      |
| 403         | `PermissionDeniedError`    |
| 404         | `NotFoundError`            |
| 422         | `UnprocessableEntityError` |
| 429         | `RateLimitError`           |
| >=500       | `InternalServerError`      |
| N/A         | `APIConnectionError`       |

## Request IDs

> For more information on debugging requests, see [these docs](https://platform.openai.com/docs/api-reference/debugging-requests)

All object responses in the SDK provide a `_request_id` property which is added from the `x-request-id` response header so that you can quickly log failing requests and report them back to OpenAI.

```python
response = await client.responses.create(
    model="gpt-5.5",
    input="Say 'this is a test'.",
)
print(response._request_id)  # req_123
```

Note that unlike other properties that use an `_` prefix, the `_request_id` property
_is_ public. Unless documented otherwise, _all_ other `_` prefix properties,
methods and modules are _private_.

> [!IMPORTANT]  
> If you need to access request IDs for failed requests you must catch the `APIStatusError` exception

```python
import openai

try:
    completion = await client.chat.completions.create(
        messages=[{"role": "user", "content": "Say this is a test"}], model="gpt-5.5"
    )
except openai.APIStatusError as exc:
    print(exc.request_id)  # req_123
    raise exc
```

## Retries

Certain errors are automatically retried 2 times by default, with a short exponential backoff.
Connection errors (for example, due to a network connectivity problem), 408 Request Timeout, 409 Conflict,
429 Rate Limit, and >=500 Internal errors are all retried by default.

Requests are retried only when their body can be safely resent.

You can use the `max_retries` option to configure or disable retry settings:

```python
from openai import OpenAI

# Configure the default for all requests:
client = OpenAI(
    # default is 2
    max_retries=0,
)

# Or, configure per-request:
client.with_options(max_retries=5).chat.completions.create(
    messages=[
        {
            "role": "user",
            "content": "How can I get the name of the current day in JavaScript?",
        }
    ],
    model="gpt-5.5",
)
```

`max_retries` must be a non-negative integer. `0` disables retries. Use a large
integer, such as `1000`, for a larger retry budget. Other values raise an error
before a request is sent.

Application exceptions raised by custom transports or hooks propagate unchanged,
including task-executor cancellation signals.

## Timeouts

By default requests time out after 10 minutes. You can configure this with a `timeout` option,
which accepts a float or an [`httpx2.Timeout`](https://httpx2.pydantic.dev/) object:

```python
import httpx2
from openai import OpenAI

# Configure the default for all requests:
client = OpenAI(
    # 20 seconds (default is 10 minutes)
    timeout=20.0,
)

# More granular control:
client = OpenAI(
    timeout=httpx2.Timeout(60.0, read=5.0, write=10.0, connect=2.0),
)

# Override per-request:
client.with_options(timeout=5.0).chat.completions.create(
    messages=[
        {
            "role": "user",
            "content": "How can I list all files in a directory using Python?",
        }
    ],
    model="gpt-5.5",
)
```

On timeout, an `APITimeoutError` is thrown.

Note that requests that time out are [retried twice by default](#retries).

## Advanced

### Logging

We use the standard library [`logging`](https://docs.python.org/3/library/logging.html) module.

You can enable logging by setting the environment variable `OPENAI_LOG` to `info`.

```shell
$ export OPENAI_LOG=info
```

Or to `debug` for more verbose logging. Set it to `warning`, `error`, or `critical`
to show only messages at that level or higher. `OPENAI_LOG` configures the `openai`
logger; configure HTTP transport loggers separately using Python logging.

### How to tell whether `None` means `null` or missing

In an API response, a field may be explicitly `null`, or missing entirely; in either case, its value is `None` in this library. You can differentiate the two cases with `.model_fields_set`:

```py
if response.my_field is None:
    if "my_field" not in response.model_fields_set:
        print('Got json like {}, without a "my_field" key present at all.')
    else:
        print('Got json like {"my_field": null}.')
```

### Accessing raw response data (e.g. headers)

The "raw" Response object can be accessed by prefixing `.with_raw_response.` to any HTTP method call, e.g.,

```py
from openai import OpenAI

client = OpenAI()
response = client.chat.completions.with_raw_response.create(
    messages=[
        {
            "role": "user",
            "content": "Say this is a test",
        }
    ],
    model="gpt-5.5",
)
print(response.headers.get("X-My-Header"))

completion = response.parse()  # get the object that `chat.completions.create()` would have returned
print(completion)
```

These methods return a [`LegacyAPIResponse`](https://github.com/openai/openai-python/tree/main/src/openai/_legacy_response.py) object. This is a legacy class as we're changing it slightly in the next major version.

For the sync client this will mostly be the same with the exception
of `content` & `text` will be methods instead of properties. In the
async client, all methods will be async.

A migration script will be provided & the migration in general should
be smooth.

#### `.with_streaming_response`

The above interface eagerly reads the full response body when you make the request, which may not always be what you want.

To stream the response body, use `.with_streaming_response` instead, which requires a context manager and only reads the response body once you call `.read()`, `.text()`, `.json()`, `.iter_bytes()`, `.iter_text()`, `.iter_lines()` or `.parse()`. In the async client, these are async methods.

As such, `.with_streaming_response` methods return a different [`APIResponse`](https://github.com/openai/openai-python/tree/main/src/openai/_response.py) object, and the async client returns an [`AsyncAPIResponse`](https://github.com/openai/openai-python/tree/main/src/openai/_response.py) object.

```python
with client.chat.completions.with_streaming_response.create(
    messages=[
        {
            "role": "user",
            "content": "Say this is a test",
        }
    ],
    model="gpt-5.5",
) as response:
    print(response.headers.get("X-My-Header"))

for line in response.iter_lines():
        print(line)
```

The context manager is required so that the response will reliably be closed.

### Making custom/undocumented requests

This library is typed for convenient access to the documented API.

If you need to access undocumented endpoints, params, or response properties, the library can still be used.

#### Undocumented endpoints

To make requests to undocumented endpoints, you can make requests using `client.get`, `client.post`, and other
http verbs. Options on the client will be respected (such as retries) when making this request.

```py
import httpx2

response = client.post(
    "/foo",
    cast_to=httpx2.Response,
    body={"my_param": True},
)

print(response.headers.get("x-foo"))
```

#### Undocumented request params

If you want to explicitly send an extra param, you can do so with the `extra_query`, `extra_body`, and `extra_headers` request
options.

#### Undocumented response properties

To access undocumented response properties, you can access the extra fields like `response.unknown_prop`. You
can also get all the extra fields on the Pydantic model as a dict with
[`response.model_extra`](https://docs.pydantic.dev/latest/api/base_model/#pydantic.BaseModel.model_extra).

### Configuring the HTTP client

You can override the [HTTPX2 client](https://httpx2.pydantic.dev/) to customize proxies, transports, authentication, event hooks, and other advanced HTTP behavior. See the [HTTPX2 migration guide](httpx2.md) when updating an existing custom client.

```python
import httpx2
from openai import OpenAI, DefaultHttpx2Client

client = OpenAI(
    # Or use the `OPENAI_BASE_URL` env var
    base_url="http://my.test.server.example.com:8083/v1",
    http_client=DefaultHttpx2Client(
        proxy="http://my.test.proxy.example.com",
        transport=httpx2.HTTPTransport(local_address="0.0.0.0"),
    ),
)
```

You can also customize the client on a per-request basis by using `with_options()`:

```python
client.with_options(http_client=DefaultHttpx2Client(...))
```

#### Mutual TLS

Before configuring a client, review the
[OpenAI Mutual TLS Beta Program](https://help.openai.com/en/articles/10876024-openai-mutual-tls-beta-program)
for enrollment, currently supported endpoints, and certificate requirements.

For API-key authenticated HTTP requests that require mutual TLS (mTLS), configure
a native [`ssl.SSLContext`](https://docs.python.org/3/library/ssl.html#ssl.SSLContext)
and pass it through the custom HTTP client:

```python
import os
import ssl

from openai import OpenAI, DefaultHttpx2Client

# Server trust is configured independently. Without `cafile`, this uses the
# operating system's normal trusted certificate authorities.
ssl_context = ssl.create_default_context(
    cafile=os.environ.get("OPENAI_MTLS_CA_BUNDLE"),
)
ssl_context.load_cert_chain(
    # This PEM must contain the leaf certificate first, followed by every
    # intermediate certificate needed to reach the server's trust anchor.
    certfile=os.environ["OPENAI_MTLS_CERTIFICATE_CHAIN"],
    keyfile=os.environ["OPENAI_MTLS_PRIVATE_KEY"],
    password=os.environ.get("OPENAI_MTLS_PRIVATE_KEY_PASSWORD"),
)

client = OpenAI(
    api_key=os.environ["OPENAI_API_KEY"],
    # A custom HTTP client does not tell the SDK that mTLS is configured, so
    # select the mTLS endpoint explicitly. Preserve an EU or custom override.
    base_url=os.environ.get(
        "OPENAI_BASE_URL",
        "https://mtls.api.openai.com/v1",
    ),
    # A client certificate belongs to the HTTP client, not the base URL.
    # Disable redirects so it cannot follow a response to another origin.
    http_client=DefaultHttpx2Client(
        verify=ssl_context,
        follow_redirects=False,
    ),
)
```

The async configuration is equivalent:

```python
import os
import ssl

from openai import AsyncOpenAI, DefaultAsyncHttpx2Client

ssl_context = ssl.create_default_context(
    cafile=os.environ.get("OPENAI_MTLS_CA_BUNDLE"),
)
ssl_context.load_cert_chain(
    certfile=os.environ["OPENAI_MTLS_CERTIFICATE_CHAIN"],
    keyfile=os.environ["OPENAI_MTLS_PRIVATE_KEY"],
    password=os.environ.get("OPENAI_MTLS_PRIVATE_KEY_PASSWORD"),
)

client = AsyncOpenAI(
    api_key=os.environ["OPENAI_API_KEY"],
    base_url=os.environ.get(
        "OPENAI_BASE_URL",
        "https://mtls.api.openai.com/v1",
    ),
    http_client=DefaultAsyncHttpx2Client(
        verify=ssl_context,
        follow_redirects=False,
    ),
)
```

See the complete [sync HTTPX2](examples/mtls_httpx2.py) and
[async HTTPX2](examples/mtls_httpx2_async.py) examples.

The certificate-bearing HTTP client is transport-wide. For API-key mTLS,
dedicate it to the selected API origin; X.509 workload identity also uses the
fixed OpenAI mTLS token-exchange origin. Do not reuse the client for unrelated
services or pass it through `with_options()` with a different `base_url`.
If redirects are required for API-key mTLS, add an
HTTPX2 request hook that rejects requests whose scheme, host, or port differs
from the configured mTLS origin before enabling `follow_redirects`.

`SSLContext.load_cert_chain()` raises during setup for unreadable or malformed
files and for a private key that does not match the leaf certificate. Certificate
expiry, key usage, extended key usage, SAN, and trust policy remain TLS server
decisions. OpenAI does not fetch missing intermediates through AIA, so provide a
complete, leaf-first client-chain PEM. Intermediate-chain support is currently
enabled by request. Until it is enabled for your organization, use a client leaf
certificate directly signed by the uploaded CA.

For certificate rotation, build a new `SSLContext`, HTTP client, and `OpenAI` or
`AsyncOpenAI` client. This creates a fresh connection pool; close the old SDK
client after its in-flight requests finish. Do not assume existing TLS
connections will renegotiate.

This recipe applies to ordinary API-key HTTP traffic. For certificate-backed
token exchange, use the X.509 workload identity configuration described above.
Realtime WebSocket mTLS is not included.

### Managing HTTP resources

By default the library closes underlying HTTP connections whenever the client is [garbage collected](https://docs.python.org/3/reference/datamodel.html#object.__del__). You can manually close the client using the `.close()` method if desired, or with a context manager that closes when exiting.

```py
from openai import OpenAI

with OpenAI() as client:
    # make requests here
    ...

# HTTP client is now closed
```

For `AsyncOpenAI`, use `async with` or `await client.close()` before shutting down
the event loop. Garbage collection cannot reliably await asynchronous cleanup.

```py
import asyncio
from openai import AsyncOpenAI

async def main() -> None:
    async with AsyncOpenAI() as client:
        response = await client.responses.create(model="gpt-5.5", input="Say this is a test")
        print(response.output_text)

asyncio.run(main())
```

## Microsoft Azure OpenAI

To use this library with [Azure OpenAI](https://learn.microsoft.com/azure/ai-services/openai/overview), use the `AzureOpenAI`
class instead of the `OpenAI` class.

> [!IMPORTANT]
> The Azure API shape differs from the core API shape which means that the static types for responses / params
> won't always be correct.

```py
from openai import AzureOpenAI

# gets the API Key from environment variable AZURE_OPENAI_API_KEY
client = AzureOpenAI(
    # https://learn.microsoft.com/azure/ai-services/openai/reference#rest-api-versioning
    api_version="2023-07-01-preview",
    # https://learn.microsoft.com/azure/cognitive-services/openai/how-to/create-resource?pivots=web-portal#create-a-resource
    azure_endpoint="https://example-endpoint.openai.azure.com",
)

completion = client.chat.completions.create(
    model="deployment-name",  # e.g. gpt-35-instant
    messages=[
        {
            "role": "user",
            "content": "How do I output all files in a directory using Python?",
        },
    ],
)
print(completion.to_json())
```

In addition to the options provided in the base `OpenAI` client, the following options are provided:

- `azure_endpoint` (or the `AZURE_OPENAI_ENDPOINT` environment variable)
- `azure_deployment`
- `api_version` (or the `OPENAI_API_VERSION` environment variable)
- `azure_ad_token` (or the `AZURE_OPENAI_AD_TOKEN` environment variable)
- `azure_ad_token_provider`

An example of using the client with Microsoft Entra ID (formerly known as Azure Active Directory) can be found [here](https://github.com/openai/openai-python/blob/main/examples/azure_ad.py).

## Amazon Bedrock

To use this library with [Amazon Bedrock's OpenAI-compatible API](https://docs.aws.amazon.com/bedrock/latest/userguide/models-api-compatibility.html), configure the standard `OpenAI` client with the Bedrock provider.

Install the optional Bedrock dependencies to use the standard AWS credential chain and SigV4 authentication:

```sh
pip install 'openai[bedrock]'
```

```py
from openai import OpenAI
from openai.providers import bedrock

# Uses your normal AWS credentials. You can omit region when it is
# configured through AWS_REGION, AWS_DEFAULT_REGION, or your AWS profile.
client = OpenAI(
    provider=bedrock(
        region="us-west-2",
    )
)

response = client.responses.create(
    model="openai.gpt-5.4",
    input="Say hello!",
)

print(response.output_text)
```

The provider configures AWS authentication and the Bedrock Mantle endpoint while retaining the normal SDK resources, retries, streaming, and error handling. AWS controls which endpoints and features are supported; unsupported calls surface the provider's normal HTTP errors through the SDK.

The default AWS credential chain supports environment credentials, shared credentials and config files, named profiles, SSO and assume-role profiles, and workload credentials such as ECS, EKS, and EC2 metadata. To select a named profile:

```py
client = OpenAI(
    provider=bedrock(
        profile="my-profile",
    )
)
```

You can also pass `access_key_id` and `secret_access_key`, with an optional `session_token`, or a refreshable `credential_provider` that returns botocore-compatible credentials. Explicit bearer and AWS credential options are mutually exclusive.

Pass `base_url` to `bedrock(...)` or set `AWS_BEDROCK_BASE_URL` to override the derived `https://bedrock-mantle.<region>.api.aws/openai/v1` endpoint. Custom URLs retain Mantle signing by default; pass `endpoint="runtime"` to use Runtime signing.

SigV4 requests require replayable, fully serialized request bodies. Standard JSON requests already meet this requirement, and response streaming is unaffected. Low-level one-shot request streams must be buffered before sending, or sent with bearer authentication and retries disabled.

Bearer tokens remain available as a compatibility or manual authentication mode. Set `AWS_BEARER_TOKEN_BEDROCK` to an [Amazon Bedrock API key](https://docs.aws.amazon.com/bedrock/latest/userguide/api-keys.html), pass `api_key`, or provide a refresh callback:

```py
client = OpenAI(
    provider=bedrock(
        region="us-west-2",
        token_provider=lambda: refresh_bedrock_token(),
    )
)
```

Without explicit authentication, `AWS_BEARER_TOKEN_BEDROCK` takes precedence over the default AWS credential chain for backwards compatibility.

### Legacy `BedrockOpenAI` client

`BedrockOpenAI` and `AsyncBedrockOpenAI` remain available for existing applications and delegate to the same provider implementation. New applications should prefer `OpenAI(provider=bedrock(...))`.

```py
from openai import BedrockOpenAI

client = BedrockOpenAI(
    aws_region="us-west-2",
    aws_profile="my-profile",
)
```

The legacy module client also continues to support `openai.api_type = "amazon-bedrock"` or `OPENAI_API_TYPE=amazon-bedrock`.

## Versioning

This package generally follows [SemVer](https://semver.org/spec/v2.0.0.html) conventions, though certain backwards-incompatible changes may be released as minor versions:

1. Changes that only affect static types, without breaking runtime behavior.
2. Changes to library internals which are technically public but not intended or documented for external use. _(Please open a GitHub issue to let us know if you are relying on such internals.)_
3. Changes that we do not expect to impact the vast majority of users in practice.

Minimum supported Python version increases are released as minor versions, not patches, when package metadata can keep users on the final compatible SDK release. See the [Python version support policy](./PYTHON_VERSION_POLICY.md) for the support window, release treatment, and compatibility history.

We take backwards-compatibility seriously and work hard to ensure you can rely on a smooth upgrade experience.

We are keen for your feedback; please open an [issue](https://www.github.com/openai/openai-python/issues) with questions, bugs, or suggestions.

### Determining the installed version

If you've upgraded to the latest version but aren't seeing any new features you were expecting then your python environment is likely still using an older version.

You can determine the version that is being used at runtime with:

```py
import openai

print(openai.__version__)
```

## Requirements

Python 3.10 or higher.

## Contributing

Please share bug reports and feature requests through [GitHub issues](https://github.com/openai/openai-python/issues).
Pull requests are limited to repository collaborators; we do not accept pull requests from non-collaborators.
See [CONTRIBUTING.md](https://github.com/openai/openai-python/blob/main/CONTRIBUTING.md) for the contribution policy and development guide.
For security vulnerabilities, follow [SECURITY.md](https://github.com/openai/openai-python/blob/main/SECURITY.md).

Repository info

Default branch: main
Latest version: v3.45.0
Created: 26/10/2020
Last activity: 05/11/2026
License: Apache-2.0
Open issues: 95
Watchers: 382

Topics

openai python

Contributors

stainless-app[bot]

704 contributions

stainless-bot

276 contributions

apcha-oai

114 contributions

jbeckwith-oai

99 contributions

RobertCraigie

96 contributions