OpenAI released ChatGPT Images 2.5, five months after 2.0. The news isn't that images come out prettier: it's that you can now touch them part by part, clicking on a specific element, and that a scribble drawn with your mouse works as valid input. The Verge tried the Sketch feature by typing @Sketch in the chat, and La Nación details the rest of the toolbar: background removal, a brush eraser, and comments anchored to a specific area of the image.
Why it matters
For two years the mental model of image generation was a slot machine: you write a prompt, pull the lever, and if you don't like it you pull again. Every iteration was a new image, not a correction of the previous one. That made the work impossible to fit into a real production process, where 90% of the time doesn't go into creating but into adjusting what already exists.
The signal hidden in this announcement is that OpenAI stopped competing on render quality and moved to competing on editing surface. Remove BG, Erase and anchored comments aren't model features: they're graphics editor features. All three solve the same problem, which is the lack of local control over a generated image.
And the context number explains it: over 3 billion images per week, according to OpenAI. At that volume, the bottleneck is no longer whether the model draws well. It's that people fire off ten generations to get one usable result. Cutting that ratio in half is worth more than any fidelity improvement, and it also explains the other reported figure: up to 50% less latency compared to 2.0.
What changes in practice
If you generate assets for products — placeholders, landing page images, mockups to hand off to design — the concrete change is that the correction cycle stops being destructive.
BEFORE (2.0)
┌──────────┐ ┌──────────┐ ┌──────────┐
│ prompt │──►│ image │──►│ no good │
└──────────┘ └──────────┘ └────┬─────┘
▲ │
└───────── rewrite ───────────┘
(the previous image is lost)
NOW (2.5)
┌──────────┐ ┌──────────┐ ┌──────────┐
│ sketch │──►│ image │──►│ comment │
└──────────┘ └──────────┘ └────┬─────┘
▲ │
└─── patch ────┘
(the rest of the composition is untouched)
Three practical consequences I see:
- The transparent PNG stops being a separate step. Remove BG exports directly. If you had a script that ran through a cutout tool before uploading the asset, that step can probably be deleted.
- The sketch replaces the long prompt for spatial composition. Describing in words where each thing goes was always text's weak point. A ten-second stroke communicates layout better than a paragraph.
- Subject consistency across images is what makes series possible. OpenAI says it improved feature preservation from reference photos and that a sketched character holds up across successive images. If that holds in real use, only then can you start thinking about a coherent set of assets instead of loose pieces.
When NOT to use it
Here's the honest part. I'll start with what the material doesn't say: there are no public benchmarks in what I was given, no API pricing, no confirmation that these editing tools are available outside the ChatGPT interface. OpenAI's announcement is short and the press coverage describes the experience in the app, not an endpoint. If your plan was to automate this from the backend, today you don't have enough information to design it.
| Case | Worth it | Why |
|---|---|---|
| Internal mockups, visual exploration | Yes | The cost of a mistake is zero and the adjustment cycle is exactly what improved |
| Final marketing assets | With human review | Localized editing speeds things up, it doesn't validate them |
| Automated pipeline via API | Wait | There's no public information about programmatic access to these tools |
| Brand identity, product typography | No | A generative editor doesn't guarantee exact reproducibility |
Two more limits worth mentioning. First: a workflow that depends on clicks and brush strokes on a canvas is, by definition, manual. It isn't versionable, it doesn't go into a repo, it can't be reviewed in a pull request. Every improvement in interactive UX is a loss of reproducibility, and in production that matters.
Second: convenience pushes volume. When editing costs little, the team generates forty variants and nobody decides which one is good. The tool accelerates production, not judgment. If your team's problem was that nobody defined the visual direction, this makes it worse.
What I'd do today
I'd try it for the one thing that's clearly solved today: replacing the back-and-forth with manual cutouts when I need a quick asset with a transparent background. That alone justifies the change and compromises nothing.
For everything else I'd wait. Before moving a pipeline I'd ask OpenAI for what it still hasn't published: API access to the editing tools, and some verifiable data on subject consistency across generations. Without that, any architecture you build on top is resting on a chat interface, and chat interfaces change without warning.