ChatGPT Images 2.5 AI image generator

Yes. ChatGPT Images 2.5 was released on September 8, 2026, and both API models run on Morphic in the browser with no download or install. Generate from a written description, or bring in up to 10 reference images and edit. Asking for ChatGPT Images 2.5 gives you GPT-Image-2.5 Flare, the fast default; name GPT-Image-2.5 Sunburst when you want the precision tier. Results land on your Canvas from any laptop, including machines with no GPU.

Morphic has a free plan, so you can sign up and start generating right away with no payment. Two things to know: downloads on the free plan carry a watermark, and ChatGPT Images 2.5 itself sits on the paid plans, alongside the other frontier models. The free plan is the right way to try the workspace and the wider catalog before deciding.

Start with GPT-Image-2.5 Flare. It matches GPT Image 2 on quality at up to half the latency, which covers most everyday work: social content, product shots, quick concepting and anything you plan to run in volume. Move to GPT-Image-2.5 Sunburst when a job needs tighter control across a chain of edits, such as production campaign creative or polished product imagery, and you can afford the longer generation time. OpenAI suggests proving the quality on Sunburst first when quality is the constraint, then testing whether Flare clears the same bar.

Output runs at 1080p, 1440p or 2160p, in nine aspect ratios from 21:9 through 9:16 plus auto, with custom dimensions supported. The model itself tops out at a 3840 pixel edge with ratios between 1:3 and 3:1. Five quality tiers are available, from low through the two new ones in this release, xhigh and max, which add fine detail for print and take longer to render.

Quote the exact words, say how many times they should appear, and describe where they sit and how they are set. Adding a line such as no extra text, no watermarks stops invented copy from creeping in. For small or dense type, compare high against xhigh before settling. Spelling unusual brand names letter by letter helps the model render them as characters rather than guess at them.

That is the improvement this release is built around. Name the single change, then list what has to stay: subject, pose, lighting, framing, background, labels. Editing precision and multi-turn stability both improved in 2.5, so a chain of revisions holds together better than it used to. If a region has to stay pixel-identical, OpenAI recommends compositing the approved edit back into the original rather than relying on the prompt alone.