GPT-Image-2.5 chooses between speed and accuracy

OpenAI is launching GPT-Image-2.5 Flare and Sunburst in its API, two models designed to speed up generation or enhance control over complex edits.

Change a jacket without moving the person, replace a background without altering the product, then correct some text without rebuilding the entire composition. With GPT-Image-2.5, OpenAI is less focused on producing another spectacular image than on avoiding undoing what already worked.

The new generation comes in two API models. GPT-Image-2.5 Flare prioritizes speed and is intended to suit most applications. GPT-Image-2.5 Sunburst takes more time but targets work that requires better overall quality and tighter control over edits.

Flare is the smaller of the two models. OpenAI recommends it for social media content, consumer creative tools, catalogs, visual exploration, prototypes, and high-volume production. The company claims a 50% reduction in latency compared with GPT-Image-2.

That promise comes with an initial caveat. The launch announcement describes higher-quality images at half the latency, while the official model selection guide more cautiously describes quality comparable to GPT-Image-2. The announced reduction is not a fixed turnaround time that applies to every request, either.

Speed depends on the prompt, the number of reference images, the requested resolution, and the selected quality level. OpenAI recommends that developers test both models on their own content rather than assume the 50% figure applies to every application.

Sunburst is the family’s flagship model. It is designed for advertising campaigns, product visuals, detailed layouts, and editing workflows where a local change must not spill over into the rest of the image. OpenAI describes its quality as superior to GPT-Image-2 but acknowledges longer generation times than Flare.

The choice, then, is not simply between a budget model and a premium version. It presents two trade-offs. Flare aims to deliver enough quality for everyday use with less waiting. Sunburst devotes more compute to precision when errors lead to further edits or make an image unusable.

The main claimed improvement concerns targeted editing. The models should better distinguish the element being changed from those that need to remain intact. A prompt can change the color of a garment while preserving the position, face, lighting, and framing. It can also replace an object, correct lettering, or change the weather without starting over with a new composition.

This capability addresses one of the recurring frustrations of generative visual creation. A seemingly simple edit can trigger a series of unintended changes. The product changes shape, the character adopts a different pose, or the typography shifts when the request concerned only a single detail.

OpenAI says Images 2.5 better understands what it should leave untouched. That distinction matters especially for editing tools aimed at photographers, brands, and studios. The quality of an edit depends not only on making the requested change successfully but also on preserving everything around it.

The improvement is also supposed to extend to successive edits. In a conversation, a user might start by replacing a background, then change an outfit, adjust the lighting, and finally revise the text. Images 2.5 is expected to retain decisions made in earlier steps for longer.

The Responses API is suited to this kind of work. It retains images and instructions within a conversation, allowing edits across multiple turns. The main model handling the conversation then calls Flare or Sunburst through the image generation tool.

The Image API remains suited to standalone operations. It provides one endpoint for creating an image from a description and another for editing existing files. Developers select `gpt-image-2.5-flare` or `gpt-image-2.5-sunburst` directly in their requests.

The two interfaces therefore serve slightly different needs. The Image API handles a clearly defined generation or edit. The Responses API supports conversational editors in which each instruction builds on the work already done.

GPT-Image-2.5 accepts text and images as input, then returns an image. It produces neither audio nor video. Image-to-video use cases involve preparing more consistent characters and settings before passing them to a separate model.

The announced improvements also cover reference photographs. Distinctive people, places, and objects should remain more recognizable when moved into a new environment or rendered in a different style. OpenAI cites better preservation of defining features, more natural lighting, and richer textures.

That fidelity does not yet amount to a persistent identity trained on a particular person. Each new request still needs the necessary references or must be part of a conversation that contains them. The documentation also acknowledges that recurring characters and brand elements can still vary across generations.

Tools designed to create a series will therefore need to retain the right references, specify exactly what must remain unchanged, and check each output. GPT-Image-2.5 reportedly reduces drift, but it does not replace a dedicated system for maintaining a person’s or product’s continuity throughout a lengthy production.

The family also improves style adherence. A prompt can define a design era, palette, technique, hierarchy, or overall art direction. The model should better preserve those characteristics as the brief becomes more detailed.

Published examples include retro posters, a wedding invitation, tourist stamps, a scientific presentation, a mosaic, and several photographic scenes. That range illustrates OpenAI’s ambition: to move beyond standalone images and produce materials with fully developed visual layouts.

Text rendering is also described as more reliable, particularly for posters, tickets, presentations, packaging, and advertisements. In theory, a brand can correct a sentence without changing its product, composition, or visual identity.

The documentation remains cautious, however. Exact text placement and legibility can still cause problems. Highly structured compositions also