People generate more than 3 billion images a week across ChatGPT and OpenAI’s API, according to the company’s own numbers. Yesterday, five months after Images 2 shipped in April, OpenAI released Images 2.5, and the update is less about a flashy new capability than about fixing the specific things that made the previous version frustrating to actually work with: edits that changed more than you asked for, consistency that degraded over a long editing session, and generation speed that made iterating feel slow.
What Actually Changed From Images 2.0
OpenAI’s own framing centers on four specific improvements rather than a single headline feature. Generation latency dropped by up to 50 percent compared to Images 2.0, a real, measurable speed gain rather than a vague “faster” claim. Image fidelity improved specifically around reference photos, meaning a photo of a real person, pet, or place you upload now carries through to the generated result more recognizably, with lighting and texture that reads as more natural rather than the slightly waxy look earlier models produced.
Editing precision is the change most likely to matter day to day. Images 2.5 is built to change only the specific element you asked for, a product, a background, a piece of text, while leaving everything else in the image untouched, even in complex scenes with multiple subjects. Multi-turn consistency addresses a related problem: in a longer editing conversation, earlier changes are now more likely to stay intact rather than subtly drifting or degrading as you layer additional edits on top of each other.
Sketch: Drawing Directly Inside ChatGPT
The standout new feature is Sketch, a tool that lets you draw directly inside ChatGPT and use that drawing as a visual reference for a final generated image. Typing “@Sketch” in a chat opens a drawing canvas where you can rough out a room layout, the silhouette of an outfit concept, or a simple doodle, then describe the style and details you want layered on top. OpenAI is explicit that this isn’t built for people who can already draw well. It’s built for anyone who finds a rough visual reference faster or clearer to produce than a fully worded text description, which covers a lot of real use cases text alone struggles with, spatial layouts especially.
Templates, Comments, and Prompt Sharing
Templates
Rather than starting from a blank prompt, new templates give you a structured starting point for popular formats, with named options like “Poster” and “Merch.” You pick a template, then fill in the specific information, design elements, and style you want, which meaningfully lowers the barrier for anyone who knows roughly what they want but struggles to describe it from scratch.
Comment-Based Editing
You can now place comments directly on specific parts of a generated image to request a focused edit, rather than describing the location and change entirely in text. For anything spatial, “make this specific object bigger” or “change the color of just this element,” pointing directly at the image is a more precise and less error-prone way to communicate the edit than a purely verbal description.
Prompt Sharing
When you share a generated image, you can now include the exact prompt that produced it, letting someone else run the same idea with their own reference photos and details. OpenAI’s own example, a viral prompt that reimagines a photo in an ’80s style, shows the intended use case directly: a shareable creative template built around a specific prompt rather than just a finished image.
The Two New API Models: Flare and Sunburst
For developers, OpenAI released two distinct models rather than a single API upgrade. GPT-Image-2.5 Flare is the default choice for most applications, delivering higher-quality output than the previous GPT-Image-2 at roughly 50 percent lower latency, and it’s positioned for high-volume, speed-sensitive use cases: social content, product experiences, visual search, and rapid prototyping.
GPT-Image-2.5 Sunburst trades speed for precision, built specifically for premium visual workflows where tighter control over edits matters more than turnaround time, production-ready campaign creative and polished product imagery being the examples OpenAI points to directly. The practical guidance from developers already testing both: choose Sunburst when editing precision is the priority, and Flare for fast, high-quality everyday generation where speed matters more than the last few percentage points of control.
What Early Partners Are Actually Saying
OpenAI shipped this release with direct testimonials from companies already building on the new API models, which is worth reading for the specific, verifiable claims embedded in the marketing language rather than the praise itself. Higgsfield AI’s head of product specifically highlighted Flare’s ability to make a meaningful edit “without losing the character, composition, or visual identity of the original image,” a direct real-world confirmation of OpenAI’s precision-editing claim from a company with no reason to overstate it if it weren’t true. Runway’s chief creative officer described the low latency and quality combination as fitting seamlessly into existing creator workflows. Manus reported Flare delivering high-quality images at two to four times the speed of the previous GPT-Image-2 in its own internal evaluations, a specific, checkable multiplier rather than a vague speed claim. Adobe is integrating the new models directly into Firefly, its creative AI studio, specifically citing faster generation and resolution consistency through repeated refinement.
Safety and Provenance: What’s Actually New Here
Images 2.5 builds on the same core safety stack as Images 2.0 rather than introducing an entirely new framework, with added safeguards specifically for risks that emerge as image models become more capable. OpenAI continues to apply C2PA metadata and invisible watermarking to help identify AI-generated images after the fact, a provenance measure that matters increasingly as AI-generated images become harder to distinguish from real photos at a glance.
Worth knowing specifically: OpenAI’s own Preparedness Framework tracks frontier risk across three categories, biological and chemical, cybersecurity, and AI self-improvement, and the company states directly that image models are not capable of the kind of code creation and execution that would enable AI self-improvement, meaning that specific risk category doesn’t meaningfully apply here the way it does to a model like GPT-6 Astra.
What This Actually Fixes for Non-Technical Users Specifically
The precision-editing improvement is easy to undersell in a feature list, but it addresses the single most common frustration non-technical users report with AI image editing generally: asking for one small change and getting back an image where several other things also shifted, a background that changed tone, a face that looked subtly different, a layout that reflowed unexpectedly. Images 2.0 and earlier models struggled specifically with isolating an edit to exactly the element requested, especially in scenes with multiple people or complex backgrounds. If Images 2.5’s own claims hold up in regular use the way early partner testing suggests, that specific failure mode, the “I asked for one thing and got five changes” problem, should show up meaningfully less often.
The multi-turn consistency fix matters for a related but distinct reason: anyone building a longer creative project inside a single chat, a series of product images, a multi-panel illustration, a campaign built from one base concept, has likely hit the problem of an image subtly degrading in quality or drifting from its original look after several rounds of edits. A model that holds earlier changes steady across more turns means fewer restarts from scratch partway through a project.
How This Fits Into the Broader AI Image Landscape
Images 2.5 arrives into a genuinely competitive field rather than an empty one. Google’s Imagen line, Adobe Firefly’s own models, and tools like Midjourney and Flux have all pushed hard on photorealism and editing control over the past year, and OpenAI is positioning this release specifically to compete on the dimensions where it had fallen behind, editing precision and multi-turn reliability, rather than trying to win on raw aesthetic style the way Midjourney has built its reputation. Adobe’s decision to integrate GPT-Image-2.5 directly into Firefly, a platform that already offers its own models, is a notable signal in itself, suggesting OpenAI’s specific gains here were compelling enough for a major creative software company to add rather than treat as a competitor to route around entirely.
That competitive context matters for anyone deciding whether to switch tools rather than assuming Images 2.5 needs to be evaluated in isolation. If your existing workflow is built around Midjourney’s aesthetic or Adobe Firefly’s licensing-clean training data, Images 2.5’s specific improvements, speed and editing precision, may not be enough reason to switch on their own. If reference-photo fidelity and multi-turn editing reliability have been your specific pain point, this release is aimed squarely at that gap.
How to Actually Start Using It
If you’re already a ChatGPT user, there’s no separate signup or opt-in required. Images 2.5 is rolling out now to ChatGPT, ChatGPT Work, and Codex users across every tier, on desktop, mobile, and web, replacing the previous image model directly in the same interface you already use. To try Sketch specifically, type “@Sketch” in any ChatGPT conversation to open the drawing canvas.
For developers, both GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst are live in the API now, with pricing detailed on OpenAI’s developer documentation site. Given the distinct positioning of the two models, default to Flare for anything volume- or speed-sensitive, and reach for Sunburst specifically when a project’s editing precision requirements justify the longer generation time.
Common Questions About ChatGPT Images 2.5
Is ChatGPT Images 2.5 available to free ChatGPT users?
Yes. OpenAI’s rollout covers all ChatGPT tiers, including the free tier, alongside ChatGPT Work and Codex, across desktop, mobile, and web.
What’s the actual difference between Flare and Sunburst?
Flare is the faster, default option built for high-volume and speed-sensitive work, delivering higher quality than the previous GPT-Image-2 at roughly half the latency. Sunburst trades speed for tighter editing precision, aimed at premium workflows like campaign creative and polished product imagery where control matters more than turnaround time.
Do I need to know how to draw to use Sketch?
No. OpenAI is explicit that Sketch isn’t built for people who can already draw well. It’s a way to communicate a visual idea, a layout, a shape, a rough composition, faster than describing it entirely in words, then let the model turn that rough reference into a finished image.
Does Images 2.5 make it harder to tell AI-generated images from real photos?
The model’s improved photorealism does make that distinction harder at a glance, which is part of why OpenAI continues applying C2PA metadata and invisible watermarking to its output, specifically to preserve a way to verify an image’s origin after the fact even as visual detection gets harder.
The Bottom Line
The most useful way to read this release isn’t as a single dramatic leap, it’s as OpenAI fixing the specific friction points that made the previous version genuinely annoying to use in real editing workflows: images that changed more than requested, consistency that broke down over a long session, and generation speed that made iteration feel slow. Combined with Sketch, a genuinely new way to communicate visual ideas that text alone handles poorly, this is a meaningfully more usable tool than its predecessor, even without a single headline capability driving the upgrade.
References and Sources
OpenAI, “Introducing ChatGPT Images 2.5”: https://openai.com/index/introducing-chatgpt-images-2-5/
OpenAI Deployment Safety, “ChatGPT Images 2.5 System Card”: https://deploymentsafety.openai.com/chatgpt-images-2-5
Axios, “OpenAI gives ChatGPT Images a major upgrade”: https://www.axios.com/2026/09/08/exclusive-hands-on-with-chatgpts-new-image-editor
9to5Mac, “OpenAI releases ChatGPT Images 2.5 with ‘sharper details’ and ‘more precise editing'”: https://9to5mac.com/2026/09/08/openai-releases-chatgpt-images-2-5-with-sharper-details-and-more-precise-editing/
Unite.AI, “OpenAI Releases ChatGPT Images 2.5 With Sketch and Two New API Models”: https://www.unite.ai/openai-releases-chatgpt-images-2-5-with-sketch-and-two-new-api-models/
About The Author
