ChatGPT Images 2.5: From Finger Sketch to Finished Art in 30 Seconds

ChatGPT Images 2.5 launched on September 8, 2026, and this time nobody is arguing about prompt tricks. Within two days, Japanese X accounts had pushed over 40,000 posts about it, explainer videos were crossing tens of thousands of views, and at least eight tech sites had published something. The release is free on every tier, which is why everyone could try it at once.

The official talking points are the usual model-marketing words: better lighting and texture reproduction, stronger subject consistency, better obedience across multiple edits. None of that explains what changed in practice. MINATO, a Japanese AI lab that runs the Planetdive note account, used the model on a real job — a birthday illustration, a product photo, a flyer — and their notes read like a better spec sheet than the announcement.

The sketch-to-finish loop

The new @Sketch command opens a canvas inside the chat. Draw something, then ask the model to clean it up. MINATO's test: a character illustration needed for a kid's birthday present, drawn with a finger on a phone. Wobbly lines, wrong proportions, embarrassing to show anyone. In about 30 seconds the model returned a colored, shaded illustration.

The detail that matters is what the model kept. The rounded outline stayed rounded, the eye that sat slightly off stayed slightly off. The cleanup happened without flattening the drawing into default-AI neatness. That is the interesting design decision. Image generation has mostly been prompt-to-ideal-image: describe, and the model renders the statistically most beautiful result. The Sketch flow flips the direction. The drawing becomes the highest-priority control signal, and the model's job is to finish what the human rough draft started.

That matters for anyone who hands generated work to a client or a print shop. A sketch carries decisions a prompt cannot: exactly this pose, exactly this expression, exactly this asymmetry. When the model preserves them, it stops being a tool that replaces your taste and becomes one that finishes it.

Partial edits that survive round three

The same session loaded a photo of a mug and asked for one change: background to white. The background turned white, the mug kept its texture, and the handle still cast the same shadow. That is the "subject consistency" claim in observable form.

Then they stacked instructions: add text in the top right, then add a discount badge. After the third instruction the mug still looked like the same object. Their note on previous models: drift usually started around the second edit. The compounding problem — every edit nudges the subject further from the original — is the real bottleneck in AI image editing, and this is one of the first releases where the chain visibly holds.

For product work this is the difference between a demo and a usable flow. Real product image jobs are never one prompt. They are a chain: swap the background, fix the price tag, correct the angle, repeat. A model that survives the chain without regenerating the product is worth more than any single-image quality jump.

The usual caveats still apply: per-plan generation caps exist, and "free" means free but not unlimited. What changed is the workflow itself.

Templates are quietly the biggest time saver

The third piece of the release is templates for flyers and product photos. Pick a template, drop in a photo, and the model places the headline, price, and discount badge. MINATO's summary is deadpan: layout work that used to take a session in Canva now ends when you insert the photo.

Templates are the least glamorous part of this release and probably the most commercially useful. Two years of AI image generation went into hero shots. The boring middle — putting a price on an image, making it look like an ad — is what actually sells. A model that produces a finished product graphic in one shot stops being a toy.

What this means for editing workflows

Three shifts worth building on:

1. Treat the sketch as a control channel. Prompt-first creation still has a place, but sketch-first lets you inject specific intent without writing a novel. A bad drawing beats a long paragraph when composition matters.
2. Think in edit chains, not generation runs. One change at a time, verify, continue. Multi-edit obedience makes this viable now, but still check each step at export size.
3. Reach for templates on repeatable assets. If you produce the same shape of thing weekly — sale cards, product shots, event flyers — build the layout once and swap the photo.

The honest read: this is not a flashy model. It is an editing model, and the jump is in the editing, not the fireworks. The direction of the conversation also flips: from "describe what you want" to "show what you want, let the model finish." That is the same control idea behind our targeted AI image editing guide and the conversational editing workflow, now with a much higher-fidelity input surface.