GPT-Image-2.5 in Production: Choosing Between Flare and Sunburst

GPT-Image-2.5 is a collection of two models: Flare and Sunburst, unveiled by OpenAI on September 8, 2026. Flare is designed for low latency and routine image generation, while Sunburst excels in editing precision, reference preservation, and higher-fidelity final output. Both models operate through the Images API for generation and editing, accepting text and image inputs and supporting quality levels ranging from auto to max.

For existing OpenAI-compatible integrations, migrating to GPT-Image-2.5 is relatively straightforward, involving changes to the base URL, credentials, and model ID. A recommended starting point is using Flare at medium quality to measure latency and image costs, then routing demanding edits or premium assets to Sunburst.

To generate an image using Flare, a server-side token must be generated and kept secure, not exposed in client-side code or logs. The request is made to the generation route, specifying the model, prompt, size, quality, and output format. The response includes the generated image data in the b64_json field, which should be decoded and stored in your system.

For editing existing assets, the /images/edits route is used. The edited image should preserve the original as much as possible, with only the specified regions or attributes modified. A mask indicating the editable region is crucial, ensuring all other pixels, text, and geometry remain intact.

When working with multiple reference images, each should be given a clear semantic role in the prompt, specifying which attributes they should influence. The Responses API can be useful when image generation is part of a larger conversational or agent workflow, allowing for a more natural integration.