News · 2026-09-08

OpenAI ships ChatGPT Images 2.5 with a drawing tool

OpenAI released ChatGPT Images 2.5 on September 8, 2026, describing it as the company's new state-of-the-art image model and pairing it with two new API models and a drawing tool. The headline engineering claim is speed: image generation latency reduced by up to 50% against Images 2.0. The headline product claim is control — a feature called Sketch that lets people draw what they want instead of describing it.

Key facts

• Latency reduced by up to 50% compared with Images 2.0.
• People already create more than 3 billion images per week across ChatGPT Images and the GPT-Image API models.
• Two new API models: GPT-Image-2.5 Flare (speed and volume) and GPT-Image-2.5 Sunburst (precision, longer generation times).
• Primary source: OpenAI's Introducing ChatGPT Images 2.5, September 8, 2026.

Most of the improvements target the same underlying frustration: the gap between the picture in your head and the picture the model returns. Three specific things changed. Reference fidelity is better, meaning that when you hand the model a photo of a person, a pet or a product and ask to see it somewhere else, the subject is more likely to remain recognisable — "image subjects look more recognizable, lighting and textures feel more natural, and distinctive features are more likely to carry through," as the announcement puts it. Editing is more surgical: ask to change one element and the model is better at changing only that element while leaving composition and background intact. And multi-turn consistency holds up longer, so a conversation with eight successive edits does not slowly degrade into mush.

That third one is the least flashy and probably the most valuable. Anyone who has actually used these tools knows the failure mode: the fourth edit silently undoes the second, or quality erodes with each round trip until you start over. OpenAI describes earlier changes as now "more likely to stay consistent, and each new edit builds on the work you've already done without degrading image quality over time."

Sketch attacks a different bottleneck. Text is a bad interface for spatial arrangement — describing "the logo slightly left of centre, above the horizon, smaller than the figure" is tedious and unreliable. Drawing it takes four seconds. Sketch lets you draw directly in ChatGPT and use that as the reference. Alongside it, OpenAI added templates for common formats like flyers and product photos, the ability to place comments directly on an image for focused edits, and prompt sharing so other people can run your idea on their own photos.

For developers, the split between Flare and Sunburst is a straightforward speed-versus-precision trade. Flare carries the quality, editing and speed improvements and is aimed at high-volume generation; Sunburst offers an extra level of precision for detailed creative work and takes longer. That is a sensible acknowledgement that a social-media pipeline generating ten thousand variants and a studio refining one hero image want different things from the same model family.

The number worth holding onto is 3 billion images a week. That is the scale at which this product already operates, and it reframes what a latency improvement means. Halving generation time at that volume is not a user-experience nicety; it is a substantial reduction in the compute bill and a change in what kinds of applications are viable — interactive editing loops become plausible where batch generation was the only option.

Why it matters, and where the caveat sits: this is an incremental product release, and it should be read as one. The announcement's claims are qualitative almost throughout — "more natural," "more reliable," "better at preserving" — with the latency figure the only hard comparative number offered. OpenAI published no head-to-head quality measurements against its own previous version or against competitors, so there is nothing here an outside party can check. That is normal for a product launch and worth naming anyway.

The larger open question is provenance, and it is not answered by a faster model. As generated images get better at preserving real people from reference photos — which is exactly what this release improves — the question of whether a given image can be identified as synthetic gets more urgent, not less. This site has tracked that gap repeatedly: Google made its visible Gemini watermark optional, Google pulled image generation from Earth a day after shipping it, and researchers found that many AI-generated images cannot be traced to any training example. A model that is markedly better at putting a recognisable real person into a new scene raises the stakes on content provenance and watermarking — a subject on which a capability release is, by itself, silent.

Primary source, verified: read the paper →

Ground Truth. (2026, September 8). OpenAI ships ChatGPT Images 2.5 with a drawing tool. Ground Truth. https://groundtruth.day/news/openai-ships-chatgpt-images-2-5-with-a-drawing-tool.html

@misc{groundtruth:openai-ships-chatgpt-images-2-5-with-a-drawing-tool,
  title  = {OpenAI ships ChatGPT Images 2.5 with a drawing tool},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/openai-ships-chatgpt-images-2-5-with-a-drawing-tool.html}
}

Topics: openai · image-generation · chatgpt · product · api