ChatGPT Images 2.5 is here: faster generation, a sketch tool, and two new API models
OpenAI's Images 2.5 replaces 2.0 across all ChatGPT tiers and the API, cutting latency by 50% and adding Sketch, comment editing, and two new model IDs.
OpenAI rolled out ChatGPT Images 2.5 on 8 September 2026, replacing Images 2.0 across every ChatGPT plan, Codex, and the API. There is no waitlist and no paid gate on the model itself. If you have used ChatGPT to generate an image since that date, you have already been using it.
The headline improvement is speed. Generation latency is down by up to 50% compared with Images 2.0, which makes the back-and-forth of iterating on a concept noticeably less frustrating. The quality improvements are real too: more natural lighting, richer textures, better preservation of subjects from reference photos, and more reliable instruction-following across multiple editing turns.
What’s new for everyday ChatGPT users
Sketch is the most interesting new feature for anyone who uses ChatGPT creatively. Type @Sketch inside a chat and you can draw directly in the app, then use that drawing as a visual reference for the image you want. OpenAI’s own examples include sketching a rough room layout or the outline of an outfit, adding a description of the style you want, and letting ChatGPT turn the rough drawing into a finished image. It is a straightforward way to communicate spatial intent without wrestling with words.
Comment-based editing lets you place comments directly on an image to target specific areas, rather than trying to describe locations in prose. You can also share your prompt so others can try the same idea with their own photos and details, which is handy for teams working on a consistent visual style.
Templates give you a starting point for common formats like flyers and product photos, so you are not staring at a blank prompt every time.
Edit history is now more useful. Each edit saves as its own file, so you can scroll back through previous versions inside Edit mode and recover a version from several changes ago. OpenAI has also improved how well earlier edits carry through longer conversations, so a change you made four turns back is less likely to quietly disappear.
What’s new for developers
The API gains two distinct model IDs, replacing gpt-image-2.
gpt-image-2.5-flare is the default choice for most applications. It is faster than its predecessor, produces higher-quality output, and handles the full range of everyday generation tasks well: social content, product imagery, visual search, rapid prototyping, and high-volume pipelines. If you were using gpt-image-2 before, Flare is the straightforward upgrade.
gpt-image-2.5-sunburst trades some speed for tighter editing control. It is built for workflows where precision matters most: production-ready campaign creative, polished product imagery, or any situation where you need to change a jacket while keeping the subject in exactly the same position, or update background copy without reworking the entire composition.
Both models accept text and image inputs and produce image output. Both are available in the Image API and as the model for the Responses API image generation tool.
Quality tiers and pricing
Images 2.0 offered low, medium, and high quality settings. The 2.5 pair adds xhigh and max above those, with auto as the new default. For a 1024×1024 image, the cost ladder runs from $0.006 at low to $0.211 at max. Token pricing is unchanged from gpt-image-2: $8 per million image input tokens ($2 cached), $30 per million image output tokens, and $5 per million text input tokens.
The short version: you get more capability at the same price, with two new quality tiers available if your use case warrants them.
A note on safety
Images 2.5 carries forward OpenAI’s existing prompt and image checks. It also adds SynthID invisible watermarking alongside C2PA metadata, so images made with the model carry provenance information that can be detected by compatible tools. Adobe’s Matt Chotin confirmed that the GPT-Image-2.5 models are coming to Firefly, which means C2PA-compatible workflows in the Adobe ecosystem will have a clear path to identifying AI-generated content from this model.
The bigger picture
OpenAI reports that more than 3 billion images are created each week across ChatGPT Images and the GPT-Image API. Images 2.5 is not a radical departure from what came before, but it is a meaningful step forward on the dimensions that affect daily use most: speed, edit reliability, and control. The Sketch feature in particular fills a genuine gap. Describing spatial relationships in text is awkward; drawing them is obvious. That it has taken this long to arrive says more about how hard the underlying problem is than about any lack of ambition.
If you are a developer, the two-model split gives you a cleaner choice than a single model with quality flags. Flare for volume and speed, Sunburst for precision. Both model IDs are live in the API from 8 September 2026.