OpenAI released gpt-image-2 two days ago, and Dawn is shipping with it today.
That matters because Dawn has been able to produce images for a while, as long as the answer was vector-shaped. Ask for a diagram, a flowchart, or a schematic and the agent draws an SVG, converts it to PNG if you need one, and drops it in the thread. That path is still the right one for anything that is fundamentally lines, boxes, and labels — it stays sharp at any size and it is easy to edit.
What Dawn could not do, until gpt-image-2, was the other half of the picture problem: illustrations, hero shots, styled icons, photograph-like renders, and edits of existing images. Those needed a different kind of model.
That model is now here, inside the same conversation.
Powered by gpt-image-2
gpt-image-2 is OpenAI’s latest image-generation model, released on 21 April 2026. It handles text-to-image generation, reference-guided variation from an existing image, and masked editing of a specific region — all of which Dawn now exposes to agents.
We wired it up the week of release, because it is a meaningful step up for the kind of visual work Dawn is asked to do day to day: product illustrations, landing-page hero shots, dashboard icons, concept renders for a new feature, and edits of pictures that already exist. Raster work is no longer a gap.
The integration uses gpt-image-2 exclusively for the raster path. Diagrams and flowcharts continue to flow through Dawn’s native SVG renderer, because that is still the better tool for that job.
SVG Still Handles the Diagrammatic Work
Nothing has changed about how the agent handles vector-native work. If you ask for a sequence diagram, a system architecture sketch, an org chart, a flowchart, or a simple graph, you still get an SVG — editable, scalable, and rasterised to PNG on request. That path is faster, cheaper, and produces better results for diagrammatic content than any raster model does.
The new capability is not a replacement for that. It is the right tool for a different class of picture.
What the Raster Path Is For
gpt-image-2 fits the things SVG never did well. Hero illustrations for a landing page, stylised icons for a dashboard, a mood image for a slide deck, a concept render for a new feature, a photo-realistic stand-in where a real photo would eventually go. You describe what you want, the agent produces it through gpt-image-2, and the image lands in the thread attached to the conversation.
You can guide the output beyond the prompt. The agent can ask for specific sizes, choose between PNG, JPEG, and WebP, toggle a transparent background when the image is going to sit on top of something else, and shift quality when a draft is enough. Most of that happens implicitly: if you ask for “a transparent logo-style icon for our dark dashboard”, the agent picks a format and background that make sense, rather than asking you to specify every knob.
It is the same agent, in the same conversation, with the same memory and context. If the thread has already been talking about your product, the image it generates tends to land closer to what you actually meant the first time.
Edit Images, Not Just Generate Them
Raster generation is only half of the release. The other half is editing.
Attach an existing image to the conversation and ask the agent to change it. Swap the background, restyle it, add an element, or produce a variation that keeps the same subject. The agent uses the attached image as a reference instead of inventing one from the prompt alone.
For more controlled edits, you can point to a specific region. If you have a mask — a second image marking the area you want changed — the agent edits only inside that region and leaves the rest of the picture alone. That is the right tool when you want to replace a product in a hero shot, change the text on a mocked-up screen, or fix a specific corner of an illustration without touching the rest.
Plain generation, reference-guided variation, and masked editing all run through the same conversation, so the agent can move between them as the work evolves.
Where Images Show Up
Generated images are attached to the conversation like any other file, which means they render inline on the channels Dawn already supports for attachments: Slack, Discord, Telegram, the web widget, Dawn Web, and Dawn Desktop. You see the image in the thread, you can download it, and you can hand it to a teammate without leaving the channel.
Teams does not currently support image attachments in Dawn conversations, so raster generation is disabled there. Teammates on Teams still see the rest of the agent’s response; the image step is the part that does not apply. If you need image work on Teams today, diagrams and flowcharts — which flow through the SVG path — still work.
Paired with HTML artifacts, the combination starts to matter. Ask the agent for a landing page mockup, then ask it to generate the hero image the page needs, and both pieces live in the same thread — ready to review together.
How Agents Decide Which Path to Use
You do not pick between raster and vector. The agent decides based on what you asked for.
“Sketch the data flow between the API gateway and the billing service” stays vector — an SVG diagram. “Make me a banner for the incident dashboard” goes raster. “Turn this screenshot into a transparent-background asset” is a raster edit. Anything that only needs text stays text.
That keeps the feature out of the way when it does not apply, and it keeps the right tool on the right job without you having to specify the format every time.
Scope Worth Knowing
A few limits are worth being up front about. The raster model returns one image per request, so asking for “four variations in a grid” will still produce a single image laid out as a grid, not four separate files. Editing requires a reference image attached in the conversation, and masked editing requires the mask to be an image too. And raster generation is not part of memory-page workflows — those stay text-and-vector-only, on purpose, so scheduled background jobs do not quietly burn through raster-generation credits.
Everything else works the way the rest of Dawn does. Images live in the conversation, they are visible to the people who can see the conversation, and they follow the same workspace and channel scoping as any other agent output.
Getting Started
There is nothing to set up. Ask an agent on any supported channel for an illustration, a hero image, a styled icon, or an edit of an existing picture, and you should get one back. Ask for a diagram or a flowchart and you still get the SVG you always did.
Before gpt-image-2 landed, a conversation that needed a photograph-like image usually ended in words about the image or a link to another tool. Now that kind of picture can sit in the same thread as the diagram next to it — rendered by whichever path fits the job.