Introducing ChatGPT Images 2.5
Simon Willison’s Weblog Simon Willison ● Covered by 4 sources
OpenAI says its image models have made more than 3 billion images. ChatGPT Images 2.5 is faster, follows instructions better, and keeps reference subjects more intact.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has a new image release, and the headline is less about flash than control. ChatGPT Images 2.5 is supposed to do three things better: follow instructions across multiple turns, return results faster, and preserve the subject in a reference photo more reliably than before.
That matters because image tools live or die on the annoying details. If you ask for a specific edit, you want the model to keep the person, object, or scene recognizable while changing the right part. OpenAI says its image models have already been used for more than 3 billion images across ChatGPT Images and the GPT-Image models in the API, so even small quality gains are likely to be felt quickly.
The API side now has two model IDs: gpt-image-2.5-sunburst and gpt-image-2.5-flare. Simon Willison’s read is pretty clear: Sunburst is the one to reach for when editing precision matters most, while Flare is the better pick for fast, high-quality everyday generation.
He also updated his openai_image.py CLI tool so it can take one or more reference images. In his example, he runs the tool with a chart image and the prompt to add a raccoon scientist studying it thoughtfully, using gpt-image-2.5-sunburst. That’s the real test here: not whether the model can make a pretty picture, but whether it can obey a messy instruction without losing the plot.
My take — AI-written commentary, not fact-checked reporting
This is the boringly important kind of AI progress: less “wow,” more “finally.” OpenAI keeps pushing image quality by tightening instruction-following and reference handling, which is exactly where these systems earn their keep. The hype crowd will chase novelty, but the people shipping tools care about control, and Sunburst sounds like the one built for that.
Read more about this at: Simon Willison’s Weblog