For three years, talking to an image model has felt like ordering coffee through a mail slot. You slide a paragraph in. Something slides back. If it’s wrong, you write a longer paragraph and try again, hoping the words carry the geometry that was in your head all along. On September 8, 2026, OpenAI cut a window into the slot. The window is called Sketch, and it matters more to agent builders than the pixel-quality gains sitting next to it in the announcement.
ChatGPT Images 2.5 shipped that day as OpenAI’s new state-of-the-art image model, with two companion models added to the API: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare carries the same quality, editing, and speed gains and is the default for most applications. Sunburst trades default status for additional precision on detailed editing. Generation latency drops by up to 50% compared with Images 2.0. Sketch lets users draw directly inside ChatGPT, invoked with @Sketch, and hand that drawing to the model as a reference. Templates for popular formats round out the release.
Read that list as a product update and it’s a solid iteration. Read it as an interface specification and something else shows up.
Latency is not a comfort feature
Halving generation time reads like a convenience improvement, the kind of thing that shortens a progress bar. Inside an agent loop, it’s structural. An agent that generates, evaluates, and regenerates is running a search over a space of candidate outputs, and the cost of each step sets the depth of search you can afford. Cut per-step cost in half and you either double the number of refinement passes inside the same budget or you move a workflow from batch to interactive.
That threshold is where agent designs change shape. Slow generation pushes you toward one-shot prompting with elaborate up-front instructions, because retries are expensive and the user is watching a spinner. Fast generation makes iterative critique loops viable: generate, run a verifier, adjust, generate again. The second pattern is more forgiving of imperfect prompts, which is convenient, because prompts are always imperfect.
Sketch adds a channel that text never had
Natural language is a terrible medium for spatial specification. Describing that the logo sits slightly left of center, overlapping the top third of the figure’s shoulder, takes a sentence that the model will interpret probabilistically. A drawing states it directly. Sketch introduces a second input channel with different properties from the first: coarse on semantics, precise on layout.
For anyone building agents on top of this API, that’s a grounding signal, and grounding signals are what reduce the ambiguity that agents accumulate over multi-turn interactions. A sketch is also machine-generatable. Nothing about a rough spatial reference requires a human hand. An agent that maintains a scene representation can render that representation as a layout guide and pass it forward, closing a loop that text alone leaves open.
The release also names identity drift across edits as one of the problems being targeted, the failure where a subject morphs between successive edits. That failure is exactly what breaks long-horizon visual agents. If step four silently changes the character’s face, every downstream step inherits the error, and there’s no cheap way to detect it without a separate consistency check. Reducing drift at the model layer removes a class of verification scaffolding that builders currently have to write themselves.
Two models is a routing problem
Flare as default, Sunburst for precision editing, is a tiered offering, and tiered offerings hand a decision to the orchestration layer. The naive approach picks one and moves on. The better approach classifies each request and routes it: bulk generation and first drafts to Flare, fine-grained edits where a small error compounds to Sunburst.
Routing well requires knowing something about the task before you run it, which means an agent needs a task classifier or an explicit signal from its planner. This is the same pattern that showed up in text models when reasoning tiers arrived, and most systems handled it badly at first by defaulting to the expensive option. The habit worth building is treating model selection as a per-call decision with a cost model attached, not a configuration constant set once at deploy time.
What to watch in your own stack
OpenAI says more than 3 billion images have been created through its tools, which tells you image generation stopped being a novelty some time ago and became infrastructure. Infrastructure gets judged on different terms than demos: predictability, consistency across calls, cost per unit of useful output.
The interesting work now sits above the model. Faster generation invites deeper loops. A drawing channel invites spatial state that persists across turns. Two model tiers invite routing logic. None of those are features OpenAI shipped. They’re openings it left, and they belong to whoever builds the layer that uses them well.
đź•’ Published: