GPT Image 2 API (October 2026): Pricing per Image, GPT Image 2.5 Flare vs Sunburst, Python Examples and Parameters

Prices and parameters below are from OpenAI's pricing page, image generation guide and model pages on October 17, 2026, unchanged since October 15. We ran the Python examples against a test endpoint to confirm the request format.

Which model to use

OpenAI's image generation guide now leads with the two 2.5 models (snapshots dated September 8, 2026) and recommends them for new integrations. gpt-image-1.5, gpt-image-1 and gpt-image-1-mini still work but are older.

Pricing

Token rates

All three models share the same rates, per million tokens:

Image output is almost all of the cost. Prompts are cheap, and reference images for edits are billed as image input. Cached input pricing only applies to the image generation tool in the Responses API, not to direct Images API calls, including edits.

Cost per image for GPT Image 2

OpenAI publishes these output costs per image for gpt-image-2:

Two things that surprise people:

• Portrait and landscape are slightly cheaper than square at the same quality. OpenAI notes that a larger non-square size can use fewer output tokens than a smaller square one.
• High quality costs 30 to 35 times low quality. Draft at low, then render the final image at high.

For the 2.5 models, OpenAI provides an interactive token calculator instead of a fixed table. They use the same token rates, but "equal token rates don't mean equal cost per image", so read usage in the response to see what each image actually cost. Streaming previews (partial_images) add 100 image output tokens per preview.

GPT Image 2 cost calculator

Image output cost only, at OpenAI's October 2026 per-image prices for gpt-image-2. Add a little for prompt and reference-image input.

Generate an image in Python

Install the SDK with pip install openai and set OPENAI_API_KEY.

import base64
from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY
result = client.images.generate(
    model="gpt-image-2.5-flare",
    prompt="An orange cat watching the rain from a windowsill, 35mm film look",
    size="1024x1024",
    quality="medium",
)
with open("cat.png", "wb") as f:
    f.write(base64.b64decode(result.data[0].b64_json))
print(result.usage)  # tokens actually used for this image

With curl:

curl https://api.openai.com/v1/images/generations \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2", "prompt": "An orange cat watching the rain", "size": "1024x1024", "quality": "low"}'

Two practical points:

• The Images API returns base64, not a URL. Decode and save it, as above.
• Complex prompts can take up to two minutes. Set your HTTP client timeout above that, and check that any proxy in between doesn't cut requests off sooner.

Edit an image

Send the source image, and optionally a mask, to /v1/images/edits (client.images.edit(...) in Python) with an instruction. The image and mask must be the same format and size, and under 50 MB. For several rounds of edits on the same image, use the Responses API with the image_generation tool, which supports multi-turn editing and file IDs as inputs. In that case you pick a regular model at the top level and set gpt-image-2.5-sunburst or gpt-image-2.5-flare in the tool's model field.

Parameters

Custom sizes on the 2.5 models: use a WIDTHxHEIGHT string such as 1536x864. Both dimensions must be multiples of 16, the aspect ratio between 1:3 and 3:1, neither edge over 3,840 pixels, and the total between 655,360 and 8,294,400 pixels (4K).

Prompting within the model's limits

OpenAI lists a few limitations worth designing around:

• Text in images is much better than before but can still miss on small or precisely placed text. Quote the exact words and keep them short.
• Recurring characters or brand elements can drift between separate generations. Edit one image instead of regenerating.
• Precise layouts aren't always followed. Expect to iterate on structured compositions.

Before you start

• Organization verification: OpenAI says you may need to complete API Organization Verification before using GPT Image models.
• No free allowance: usage is billed from prepaid API credits. See How to get an OpenAI API key.
• Supported countries: OpenAI warns that using the API outside its supported countries and territories may get your account blocked.

Errors

Using GPT Image through a gateway

Gateways that support OpenAI's Images API work with the same code: change base_url and the key. api.foo, the service that publishes these docs, serves GPT Image models at https://api.foo/v1/images/generations. We confirmed the path with a test key using the Python example above. Its image models are billed per image rather than per token; model names and prices are on the api.foo pricing page. For long renders, its direct route https://dc.api.foo/v1 has no 100-second proxy timeout.

Related

• Nano Banana API: Gemini image generation
• Grok API, including Grok Imagine
• OpenAI API pricing and calculator
• How to get an OpenAI API key

Sources

OpenAI documentation, checked October 17, 2026: Image generation guide, Pricing, model pages for gpt-image-2, gpt-image-2.5-flare and gpt-image-2.5-sunburst, the Images API reference, Supported countries and Organization verification.