OpenAI Releases ChatGPT Images 2.5, Cutting Image Generation Time in Half

OpenAI announced the launch of ChatGPT Images 2.5 on September 8, 2026, the latest version of its image generator built into ChatGPT. The company says the new model produces images up to 50 percent faster than the previous 2.0 version from April, while doing a better job of preserving faces and details across repeated edits of the same image.

The new version rolled out simultaneously on desktop, mobile, and in the browser, covering every ChatGPT subscription tier, including the free plan. OpenAI notes that usage of its image generator is already substantial, with the number of images created weekly in ChatGPT and through the API surpassing 3 billion.

What actually changed

The most visible addition is the Sketch tool, which lets users draw a simple sketch with a finger or stylus on a touchscreen that the model then uses as a reference point for the generated image. This is joined by ready-made templates for popular formats such as flyers or product photos, the ability to leave comments directly on parts of an image, and the option to share prompts with other users for collaborative work on a project.

OpenAI highlights improvements in overall image quality, describing the outputs as having more natural lighting and richer textures than version 2.0. The model is also said to better understand complex, multi-step instructions, so users no longer need to re-describe every element of a composition each time.

Faces, animals and targeted edits

A key change concerns what happens during successive edits of an already generated photo. Earlier versions of the generator could, when making a small tweak such as changing the lighting on a background wall, end up redrawing an entire person's face in the process. ChatGPT Images 2.5 is designed to make truly targeted changes instead, preserving the identity of faces and animals across successive versions of the same image.

Two models for developers

Developers using the OpenAI API now have two new models to choose from. GPT-Image-2.5 Flare is the default variant, optimized for speed and high-volume generation, with better support for transparent backgrounds, which is useful for creating brand assets. GPT-Image-2.5 Sunburst runs slower but offers the higher precision needed for tasks like product photography, where every pixel matters.

API pricing remains unchanged despite the performance gains. Input text costs $5 per million tokens, or $1.25 with cached input, input images cost $8 per million tokens, and generated output images cost $30 per million output tokens. Text generated by the model is not billed separately.

Safeguards and content labeling

OpenAI states that the new version retains its prompt and output moderation systems and continues to tag generated images with C2PA metadata and an invisible watermark, intended to make unauthorized use of the images and the creation of deepfakes harder. This matters given the growing number of disputes over the provenance and use of AI-generated content playing out simultaneously in many countries.

For users in Poland, the change mainly means that ChatGPT's advanced image generator is now available at no extra cost, including on free accounts, rather than being limited to paid plans. For businesses and creators working with advertising or product imagery, the ability to choose between the fast Flare model and the more precise Sunburst model, depending on whether speed or final file quality matters more, could prove decisive.

The image generator industry is moving very fast this year, with OpenAI competing in this space against, among others, Google's Nano Banana Pro models. The pace of updates, roughly every few months, shows that for major players, the quality and speed of image generation have become one of the main battlegrounds alongside language models themselves.