OpenAI rolls out ChatGPT Images 2.5 with faster generation, improved editing

OpenAI on Tuesday, Sept. 8 launched ChatGPT Images 2.5, an updated image-generation model that the company says improves image quality, editing accuracy and generation speed.

The update is being rolled out to ChatGPT, ChatGPT Work and Codex users on desktop, mobile and web. OpenAI is also making two new models like GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst available through its Application Programming Interface (API) for developers.

What has changed

OpenAI said Images 2.5 produces more natural lighting, richer textures and sharper details, while improving its ability to preserve subjects from reference photographs. It also aims to follow editing instructions more consistently when users make several changes to the same image.

The company said it has reduced image-generation latency by up to 50% compared with Images 2.0. The stated improvement is intended to allow users to generate and revise images more quickly.

Another focus is targeted editing. OpenAI said users can make changes to individual elements of an image while preserving other details, including the subject, composition and surrounding background.

The model is also designed to handle more complex layouts and visual instructions, including images containing real-world information, transparent backgrounds and defined visual styles. OpenAI’s system card describes improved consistency during image editing and better accuracy in infographics and layouts.

New tools for ChatGPT users

Alongside the model update, OpenAI has added several features aimed at making image creation less dependent on detailed text prompts.

One of them, Sketch, allows users to draw directly in ChatGPT and use the drawing as a reference for the generated image. Users can also start with templates for formats such as posters and merchandise, add comments directly to generated images and share the prompts used to create them.

The Verge, which tested the Sketch feature, reported that a rough drawing could be turned into a detailed AI-generated image. It also tested image annotations for requesting specific changes.

The additions reflect a broader shift in AI image tools toward editing and iteration, rather than generation from a single prompt. For users, this means an image can be created and then modified through a sequence of instructions without starting over each time.

Derya Unutmaz, who has experimented with AI image generation since the early Midjourney era, described Images 2.5 as a significant improvement over its predecessor. “It’s a major upgrade to the Image 2.0 model, which was already the best image model for me!” he wrote, while also praising the model’s ability to produce detailed AI art paintings. 

Companies begin integrating the models

OpenAI said several companies are among the early users of its new API models.

Adobe has added GPT-Image-2.5 models to Firefly. Matt Chotin, senior director of product at Adobe, said the integration gives creators another model option while maintaining image quality through repeated refinements.

Runway Chief Creative Officer Jamie Umpherson highlighted the combination of image quality and low latency, while Higgsfield AI Head of Product Axultan Alimkulov said the model was particularly effective at determining what should remain unchanged during an edit.

Manus, another early customer, said its internal evaluations found that GPT-Image-2.5 Flare generated high-quality images at two to four times the speed of GPT-Image-2. That figure is also an assessment supplied by the company and published by OpenAI.

Two models for developers

For developers, OpenAI is introducing GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.

Flare is positioned as the faster option and is intended for applications such as social content, product experiences, visual search, prototyping and high-volume image generation. OpenAI said it provides higher-quality images than GPT-Image-2 at 50% lower latency.

Sunburst is aimed at more detailed creative and editing workflows where greater control is required, with longer generation times.

The API update also adds additional image-generation options, including more image sizes, higher quality settings and background controls.

Safety and provenance

OpenAI said Images 2.5 uses checks on prompts and images to prevent harmful outputs. The company also continues to use Coalition for Content Provenance and Authenticity (C2PA) metadata and invisible watermarking to help identify images generated with its systems.

Its system card outlines evaluations covering safety risks, image provenance and safeguards used during deployment.

The launch comes as AI image-generation companies increasingly compete on editing accuracy, consistency and speed alongside visual quality. With Images 2.5, OpenAI is placing particular emphasis on the ability to preserve an existing image while making specific changes, a capability that could be relevant to creative, marketing, product and media workflows.

Mansi Hake