There are two doors to the same models. The Images API is the direct one: prompt in, image out. The Responses API lets a chat model decide to call an image_generation tool as part of a conversation, so you can refine an image in follow-up turns.
1. Key
export OPENAI_API_KEY="sk-..." # macOS/Linux; PowerShell: $env:OPENAI_API_KEY="sk-..."
pip install openai
Both SDK calls below read the key from the environment.
2. Images API: prompt in, PNG out
from openai import OpenAI
import base64
client = OpenAI()
result = client.images.generate(
model="gpt-image-2",
prompt="A children's book drawing of a veterinarian using a stethoscope to listen to the heartbeat of a baby otter.",
)
with open("otter.png", "wb") as f:
f.write(base64.b64decode(result.data[0].b64_json))
Returns the image as base64 in data[0].b64_json; decoding it and writing the bytes gives you a PNG. gpt-image-2 is the current model; gpt-image-1.5, gpt-image-1 and gpt-image-1-mini are the cheaper/older options.
3. Responses API: the model generates as a tool
from openai import OpenAI
import base64
client = OpenAI()
response = client.responses.create(
model="gpt-6-astra",
input="Generate an image of a gray tabby cat hugging an otter with an orange scarf",
tools=[{"type": "image_generation"}],
)
images = [o.result for o in response.output if o.type == "image_generation_call"]
if images:
with open("otter.png", "wb") as f:
f.write(base64.b64decode(images[0]))
The text model plans the prompt and calls the tool; the image comes back as an image_generation_call output item with base64 in result. Continue the same conversation (“make the scarf blue”) to edit it.
4. The knobs
result = client.images.generate(
model="gpt-image-2",
prompt="Flat vector icon of a paper plane, single color, transparent background",
size="1024x1024", # or 1536x1024, 1024x1536, auto
quality="high", # low, medium, high, auto
output_format="png", # png (default), jpeg, webp
background="transparent" # transparent, opaque, auto
)
n makes several images per call; output_compression (0–100) shrinks JPEG/WebP; moderation="low" relaxes the content filter — blocked requests come back as a moderation_blocked error.
5. Edits
The edits endpoint takes an input image (and optionally a mask marking the region to replace) plus a prompt, and returns the modified image in the same base64 shape. It’s the route for “same picture, change one thing”; for a new picture, generate again.
Next: your first OpenAI API request · generate images locally with FLUX.