Back to changelog

v2026.6.1: gpt-image-2 image editing fixed

gpt-image-2 image-to-image edits now work end to end — supplying an input image no longer returns HTTP 500. The gateway routes image-edit requests to OpenAI's multipart edits endpoint. Text-to-image generation is unchanged. The unstable osmQwopus3.6-27B model has been retired.

gpt-image-2 image editing fixed

Sending gpt-image-2 a request that included an input image (to edit an existing picture instead of generating from text) failed with HTTP 500 on every retry. Text-to-image generation was never affected.

Root cause

OpenAI serves text-to-image and image-editing through two different endpoints:

  • POST /v1/images/generations — JSON body, text prompt only.
  • POST /v1/images/edits — multipart/form-data, source image(s) uploaded as files.

The gateway was sending edit requests (those with an image_url in the message) to the generations endpoint with the image attached as a JSON field. OpenAI rejected this with 400 Unknown parameter: 'image', which the gateway surfaced to callers as a 500.

The fix

When a gpt-image-2 request includes one or more input images, the gateway now routes it to /v1/images/edits and uploads each image as a multipart file (image[]), exactly as OpenAI's API expects. Both inline data: URLs and remote https:// image URLs are supported, and multiple input images can be sent. The edited image is returned in the same base64 format as before — no caller changes required.

Tip: include one or more image_url parts alongside your text prompt to edit an existing image. See the image generation docs for a full example.

Retired model — osmQwopus3.6-27B

osmQwopus3.6-27B-mlxmtp-oQ8 (Neysa-hosted) has been removed from the catalog after a sustained period of upstream errors. Requests for this model id will no longer route. No other models are affected.

Compatibility

No caller changes required. The image-edit fix is scoped to OpenAI gpt-image-* requests that carry input images; text-to-image generation and all other image providers (Google Imagen, Gemini chat-image, Alibaba qwen-image-*, Z.AI cogview-*) are untouched.