OpenAI launches two new image models with separate API tracks

OpenAI has released a pair of image-generation models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, and split access between two distinct API surfaces built for different workflows. According to the official documentation, Flare targets everyday, high-throughput image creation, while Sunburst is aimed at workflows where editing precision is the primary concern. Developers must select the appropriate model on their very first call.

The Image API exposes two classic endpoints — Generations for creating from scratch and Edits for modifying an existing image — with the model name specified directly in the model parameter. The Responses API, by contrast, treats image generation as a built-in tool inside a multi-turn conversation. It accepts file identifiers rather than only raw bytes and lets a primary language model, such as GPT-6-Astra, decide when to produce a new image and when to edit the previous one. The choice between APIs follows the application architecture: a single image from a single request goes to the Image API; a conversational, iterative experience goes to the Responses API.

Both paths share the same output controls — quality, size, format and compression — and the n parameter for returning multiple images in one request, defaulting to one. The documentation illustrates a synchronous Image API call that decodes Base64 and writes to disk, versus an asynchronous Responses API flow where the result arrives inside an output object of type image_generation_call.

The Responses API's standout capability is multi-turn editing. Callers can feed the previous output back via previous_response_id or the image's own identifier and continue refining the visual with additional text prompts. An action parameter on the image_generation tool governs behavior: "auto" leaves the decision to the model, "generate" forces a fresh creation, and "edit" forces modification of the existing image in context.