QWE AI Academy How do you get ChatGPT Images 2.5 to change only the part you care about – without rewriting the whole prompt every time?
That question is the whole product, not the “50% faster” headline. Images 2.5 rolled out September 8, 2026. Below: the clicks, the control surfaces, and three gotchas announcement posts usually skip.
What actually changed for daily use
Latency drops by up to 50% versus Images 2.0 (OpenAI’s launch numbers). Lighting looks less plastic; textures hold more detail; faces and pets from reference photos stay closer to the source. Multi-turn edits hold up better than 2.0 when you stack small changes. You can keep chatting while a heavy job runs – help still warns some requests take a few minutes.
Control is the part that matters day to day. OpenAI’s launch post ships Sketch, format templates, on-image comments, and prompt sharing. On the API side you pick GPT-Image-2.5 Flare (fast default) or GPT-Image-2.5 Sunburst (tighter edits, longer waits). Image gen sits on every ChatGPT tier; “Images with thinking” stays on Plus, Pro, and Business for now (Enterprise/Edu called out as coming soon in the same docs).
Step-by-step: generate, sketch, pin
Paths match the current Images in ChatGPT help article as of the 2.5 rollout.
- Basic generate – Open a chat (web, desktop, iOS, or Android). Describe the image in plain sentences, or sidebar → Images → prompt. Optional: aspect ratio in the picker or written into the prompt.
- Templates – Sidebar → Images → Templates. Category (Poster, Merch, product photo, flyer…), answer follow-ups for text/style/layout, upload a reference if offered, send. Spell out the copy and the job (“Instagram square product hero,” not “nice poster”).
- Sketch – Type
@Sketch(or@→ Sketch). Block in layout or silhouette. Checkmark on mobile. Add a short style line – “matte ceramic mug, soft left light, plain desk.” Rough shapes beat pretty doodles; the model reads structure first. - Targeted edit – Open the image → Edit. Select a region or drop a comment pin (“eyes blue,” “kill the logo top-right”). Chat-only edits still work with no selection. Remove BG / Erase when you need a cutout or one object gone (early testing + help editor tools).
- Save / share – Autosave under Images. Download, copy, or Share; ship a prompt template when others should remix with their own photos. Delete the conversation if you need the image gone from history.
Pro tip: Treat Sketch as a composition lock. One clear silhouette + three style words beats a gallery-ready doodle.
After a product-photo template, try: “Keep the bottle label readable, soft studio light from the left, matte black background, slight surface reflection, square crop for Instagram.”
Common pitfalls that burn generations
The catch is selection bleed. OpenAI’s help center states it flatly: highlights aren’t always precise, and edits can run past the area you marked. Brush larger than the object, or pin a comment that names it – “the blue mug only,” not “here.”
Templates still aren’t in Work mode. Work account and no Templates row? Plain prompts or a personal chat.
Free-tier image caps exist; OpenAI does not publish a stable daily number. You hit a wall, then wait or upgrade. Plus/Pro raise the ceiling (Pro often treated as high/”unlimited” under abuse guardrails). “Up to 50% faster” is a ceiling on latency vs 2.0 – not a stopwatch guarantee on dense layouts that still “may take a few minutes.”
When to stay vs switch tools
Skip the leaderboard screenshots; match the tool to the job.
| Need | Lean toward | Why |
|---|---|---|
| Iterate inside a chat you already use | ChatGPT Images 2.5 | Sketch + comment pins + multi-turn memory in one thread |
| Max stylized “look” exploration | Midjourney | Still the default for many pure art-direction grids |
| Google stack + quick free trials | Gemini image tools | Lowest friction if you already live there |
| API volume vs precision stills | Flare vs Sunburst | Flare for speed/volume; Sunburst when one element moves and the rest must freeze |
Mixed writing + research + visuals in one place? ChatGPT. Vibe-only variation grids? A dedicated art tool still wins.
One open question nobody’s settled: how often multi-turn fidelity holds after 8-10 stacked micro-edits on a crowded scene. Early use looks cleaner than 2.0; long chains still deserve a fresh branch when quality softens.
FAQ
Is ChatGPT Images 2.5 free?
Yes – every tier, with caps. Free hits an unpublished daily wall. Paid plans get more headroom; thinking-mode image work stays on Plus/Pro/Business for now.
How do I use Sketch on desktop vs phone?
Same entry everywhere: @Sketch in the box (or @ → Sketch). Mobile help flow is draw → colors/erase/undo → checkmark → short text style. Desktop uses the same trigger; mouse lines stay ugly and that’s fine. After render one, pin comments instead of redrawing the whole frame.
Flare or Sunburst if I’m building an app?
Default to Flare. Higher quality than prior GPT-Image-2 at roughly half the latency – social, prototypes, volume jobs. Sunburst only when a production still needs one element nudged and everything else locked; you pay in wait time. As of launch, Flare docs list ballpark token rates around $5/1M text in, $8/1M image in, $30/1M image out – confirm live pricing before you ship. Both take text + image in and return images.
Open ChatGPT. Type @Sketch. Draw the ugliest layout of the thing you need today. Force one comment-pin edit before you quit. That’s the upgrade in five minutes.