AI&Design
Back to timeline

ChatGPT image · Announcement · Official source

ChatGPT Images 2.6 in practice

On November 4, 2026, OpenAI introduced Images 2.6 across ChatGPT tiers and said the model generates faster and better preserves material such as a character's outfit during a background change. The launch announcement gives a 50% faster generation figure and separately describes selection of Sunburst and Flare in the API, including a six-second best case for low-quality Flare. The September Images 2.5 release had already introduced image-editing controls in ChatGPT. November brought the existing markup workflow to the API. ChatGPT release notes · November launch announcement

What operators saw in ChatGPT

On November 5, AInformation used ChatGPT Plus on the web with thinking amount High to create a penguin mascot, then asked in the same thread to replace the white background with a snowy shopping street. One edited still appeared after 51.3 seconds. The tester reported that the yellow raincoat, red scarf, blue boots and round glasses held up. The published still screenshot shows only the upper portion, so the tester used subsequent frames to check the full body. A separate sign generation took 114.3 seconds on 2.6 Plus and 89.1 seconds in an October 2.5 Pro trial. The two runs used different plans and settings. AInformation's prompts, screenshots and timings

In Analytics Vidhya's launch-week report, Vasu Deo Sankrityayan opened a room photo in ChatGPT, drew a red rectangle around an armchair with Markup and requested forest-green velvet upholstery. The published output shows a green chair, a white cushion and a similar room composition, with the red guide removed. The report covers one edit with no timing or pixel comparison. Analytics Vidhya's marked input and output

On November 11, a self-described ChatGPT Pro subscriber reported poor reference fidelity and posing in multi-character images while using Extra High thinking. The post identifies Images 2.6 and the user's thinking setting. The specific image-serving variant remains unverified. A separate November 7 post described problems with established reference workflows dating to October 28, before the November 4 announcement. The second author did not identify a model version. November 11 Pro report · November 7 workflow report

PhotoWorkout's model-selected comparisons

PhotoWorkout's November 7 trial published four before-and-after photo edits labeled maximum-quality 2.6 Sunburst. A portrait background swap preserved the sitter's recognizable features and reflections while smoothing the skin slightly. Relighting kept the house's windows, door and fence close to the original. Removing a phone and clutter also changed a binder clip and rewrote handwriting on a notebook. A storefront watercolor kept the large signs readable while small notices turned into strokes. The four edits took 39–60 seconds. PhotoWorkout's four paired edits and detail crops

For a separate still-generation timing test, PhotoWorkout repeated one 1024×1024 prompt three times per model through OpenClaw with ChatGPT subscription authentication. High-quality Sunburst averaged 30.1 seconds on 2.6 and 31.9 seconds on 2.5. Low-quality Flare averaged 31.2 seconds on 2.6 and 29.3 seconds on 2.5. Those figures describe one prompt in PhotoWorkout's pipeline. The article documents its later move into native ChatGPT for separate tests. PhotoWorkout's timing method and route

The launch-week reports show where preservation held and where it failed on specific edits. The ChatGPT examples support a practical distinction between keeping a recognizable outfit or room and preserving every small feature. The available timings leave the size of any ChatGPT-wide speed gain undetermined.