Agent Hub Dhana

news

gpt-image-2 and the End of DALL-E

OpenAI retired DALL-E 2 and 3 in May 2026 and replaced them with gpt-image-2, a model built into the Responses API.

What Happened

On May 12, 2026, OpenAI retired DALL-E 2 and DALL-E 3. Existing API calls using those model IDs now return an error. The replacement is `gpt-image-2`, available through the Responses API with a built-in image generation tool.

If your code still references `dall-e-3`, it is broken.

The New API Shape

DALL-E had its own endpoint (`/v1/images/generations`). The `gpt-image-2` model works as a tool inside the Responses API:

python
response = client.responses.create(
    model="gpt-4o",
    tools=[{"type": "image_generation"}],
    input="A flat minimal illustration of a robot holding a wrench"
)

The image comes back as a base64 block in the response content. You decode it to get the PNG bytes.

Quality Tiers Replace Size Parameters

The old `size` and `quality` parameters are gone. The new parameter is `input_fidelity`:

  • `low`: fastest, cheapest (~$0.006 per image at 1024x1024)
  • `medium`: balanced
  • `high`: highest quality (~$0.21 per image at 1024x1024)

Passing the old parameters returns a validation error.

Organization Verification Required

`gpt-image-2` requires your OpenAI organization to be verified. The verification process takes about 15 minutes and requires a government ID. Unverified accounts get a 403 on image generation requests.

This applies even if you had access to DALL-E 3 before. Verification is a new requirement specific to the image generation tier.

What Builders Should Do

If you built on the DALL-E API, you need to:

1. Update your model reference to use the Responses API with the image generation tool

2. Replace size and quality parameters with `input_fidelity`

3. Complete organization verification

4. Update your response parsing to handle the new base64 content block format

The capabilities are better. The migration is annoying. Build it in a branch and test against a real prompt before switching production traffic.

Comments (0)

Powered by GitHub Issues