ChatGPT Images 2.5 Review: What's New and Worth Knowing

On September 8, 2026, OpenAI shipped the successor to ChatGPT Images 2.0, and this time the interesting part is not the image model itself — it is the way you work with it. The standalone DALL·E GPT has been retired, so for most people the Images hub inside ChatGPT is now the only door into OpenAI's image generation. The company reports that more than 3 billion images are created every week across ChatGPT Images and the GPT-Image models in the API.

This ChatGPT Images 2.5 review is research-based rather than hands-on. It pulls from the official announcement, the help center, the system card, and the API pricing page, and it answers one question: what is materially different for the person actually creating images?

What ChatGPT Images 2.5 Actually Changes

It is tempting to read ChatGPT Images 2.5 as a sharper renderer. It is that, but reducing the update to image quality misses the point. Three separate things shipped on the same day.

First, the model itself produces more natural lighting and richer textures, and OpenAI says it renders real-world information — signs, maps, diagrams, infographic-style layouts — with better accuracy. Second, the editing behavior changed: the model is designed to change only what you ask for and to hold on to reference-photo subjects across edits. Third, the product around the model gained new entry points: Sketch, templates, comments, and prompt sharing.

For a creator, the practical effect is a shift in how a session runs. The first generation stops being the end of the task and becomes a working draft you can revise in place. That framing — reference first, then targeted edits instead of full regenerations — is the core of what the 2.5 workflow actually is.

What's New in the 2.5 Workflow

Four changes move the biggest needles in day-to-day use: how much of your source image survives, whether an edit stays local, whether a long editing conversation stays coherent, and what you can hand the model instead of a text prompt.

More Consistent Reference Images

Reference fidelity is the quietest but most valuable upgrade. OpenAI says image subjects look more recognizable and that distinctive features are more likely to carry through when you take a familiar face, pet, or product into a new setting, style, or composition. That is not a promise of pixel-identical output — faces can still drift in busy scenes — but it changes what you can reasonably attempt.

For example, a headshot can become a portrait in a new style, or a product shot can move to a new background, and the subject still reads as the same person or object. That anchors a whole class of work that previously required regenerating until the identity survived. With ChatGPT Images 2.5, keeping the subject through a transformation becomes the default expectation rather than a lucky outcome.

More Precise Editing

Editing is where OpenAI put the most emphasis in ChatGPT Images 2.5. The model is better at editing only what you asked for while keeping the rest of the composition — background, pose, lighting — intact, even with complex subjects. In the app you describe an edit in words, or highlight the region first with the selection tool and describe the change in the conversation.

The honest caveat comes from the help center: highlights are not always precise, and edits can extend beyond the area you selected. On mobile the image controls also include resize and remove tools. The practical win is that a label, a color, a background element, or a piece of text can be swapped without regenerating the whole image.

Better Multi-Turn Consistency

Earlier OpenAI image models tended to degrade as a conversation stretched on: edit number seven often had little in common with edit number one. OpenAI says ChatGPT Images 2.5 follows editing instructions more reliably across multiple turns, that earlier changes are more likely to stay consistent, and that each new edit builds on the previous work instead of slowly erasing it.

The phrase that matters is "more likely," not "always." Still, for campaign variants, storyboard characters, or marketplace-cropped product shots, staying in one conversation and keeping the thread coherent is a real workflow change rather than a spec-sheet detail.

Sketch, Templates, and Comments

Three product features change how you start and how you give feedback. Sketch is the headliner: type @Sketch in the mobile app, draw with your finger, and add a written description — the model treats your rough drawing as a visual guide, not a mask. The same idea in a wider sketch-to-image workflow happens here in the same conversation where you generate. These features are part of the same ChatGPT Images 2.5 rollout, so there is nothing separate to install.

Templates give non-designers a starting point instead of a blank canvas. In the sidebar, Images → Templates lets you pick categories such as posters, flyers, merch, product photos, or logos, customize the details, and let ChatGPT ask a couple of clarifying questions before generating. One gap worth knowing: templates are not yet available in Work mode.

Comments are a targeted-editing aid: place a note on a specific spot of an image and the next generation responds to it, instead of stacking every change into one long paragraph. Nothing in the current documentation describes shared team annotation, so treat comments as a personal editing tool today. When you share an image, you can also include the prompt that created it — Share, then Prompt template, then Copy link on mobile — so someone can reuse your layout with their own photos and details.

ChatGPT Images 2.5 vs Images 2.0

The model is not a new product; it is the next version of the image engine you already use. It appears in the same place in ChatGPT, same Images entry point, same subscription setup — no new plan is required to see it.

What actually changed versus Images 2.0:

• Generation latency lowered by up to 50%, per OpenAI. That is the headline.
• Sharper detail, more natural lighting and richer textures in output.
• Better reference-photo subject preservation, the anchor for edit-led work.
• More precise editing that keeps untouched regions stable.
• Stronger multi-turn consistency across repeated edits.
• New workflow tools — Sketch, templates, comments, prompt sharing — around the same model surface.

One of the "up to 50%" claims is worth treating carefully: it is OpenAI's stated maximum improvement, not a measured baseline guarantee, and complex requests may still take a few minutes. Early independent coverage has made the same point about the absence of a published latency methodology. If you want a broader look at how it sits against other generators, our roundup of free AI image generators and the Nano Banana vs Midjourney vs DALL-E comparison put it in context.

Flare vs Sunburst: What's the Difference?

The naming causes real confusion, so it is worth being explicit: ChatGPT Images 2.5 is the product name for image generation inside OpenAI's apps, while the API exposes two separately billed models. They share the same underlying capability but are aimed at different jobs.

GPT-Image-2.5 Flare is the default for most applications. OpenAI positions it as delivering higher-quality images than GPT-Image-2 at up to 50% lower latency, suited to creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation. In launch comments, customers such as Manus described Flare as two to four times faster than GPT-Image-2 in their own evaluations, with improved transparent-background output.

GPT-Image-2.5 Sunburst trades speed for control. It is built for premium visual workflows — production-ready campaign creative, polished product imagery, detailed edits — with longer generation times.

The decision rule is simple because pricing is not: both models bill identically today. Text input runs $5 per million tokens, image input $8 per million tokens, and image output $30 per million tokens, with cheaper cached input, per the API pricing page. That matches current GPT-Image-2 pricing, so you pick by job rather than by bill. This is API pricing and stays separate from any ChatGPT subscription you may hold.

Where ChatGPT Images 2.5 Is Most Useful

The strengths cluster around work that needs a reference, a revision, or a format — not blank-canvas art from scratch. In practical terms, ChatGPT Images 2.5 pays off most when the revision matters more than the first generation.

• Photo-driven work. Portraits, pet pictures, and product shots that must stay recognizable through new backgrounds or styles.
• Revision-heavy sessions. Campaign variants, storyboard characters, thumbnails, and marketplace-adapted product images where each change should preserve earlier ones.
• Infographics and real-world content. OpenAI highlights more accurate images that include real-world information, and the system card points to improved infographic accuracy and layout.
• Transparent backgrounds. The rollout coverage includes improved transparent-background generation for logos, brand assets, presentations, and websites.
• Non-designers. Templates handle formats, and Sketch handles layouts people cannot describe in words.
• Developers. The API path starts at Flare for most workloads and reaches for Sunburst where edit control dominates.

What It Still Does Not Solve

A review that is purely positive is not a review. The limits below are grounded in what OpenAI documents or what independent early accounts have reported.

• "Up to 50% faster" is a ceiling, not a baseline. OpenAI has not published a latency methodology, and complex requests can still take a few minutes.
• Selection highlights can bleed. The help center is explicit that highlights are not always precise.
• Reference fidelity is improved, not guaranteed. Faces and products can still drift, especially in busy compositions.
• Templates are missing from Work mode for now, and sketch drawing is described on the mobile app.
• Long text can still slip. Independent early hands-on accounts say short labels render cleanly more often, but dense passages can still render with errors. If you need long copy inside the visual, keep the copy in the layout.
• The API is per-token. It is a fit for products and pipelines, not a cheap stand-in for casual subscription generation.
• No pixel-level masking. The selection tool and comments are closer to guided ambiguity than the layer control of a pro raster editor.

The honest summary is that ChatGPT Images 2.5 solves workflow problems well, while specialist tasks still need their specialist tools.

Safety, Provenance, and Privacy

OpenAI keeps the safety stack from Images 2.0 and adds to it. The system card documents upstream refusals that screen requests before generation, and a safety reasoning model that checks both input images and final output, blocking violative content at either layer. Live blocking and offline review of flagged conversations round out the enforcement, and the card notes OpenAI may suspend accounts where ongoing misuse is detected.

The same card is candid about a new risk: the heightened realism of ChatGPT Images 2.5 could, absent safeguards, enable more convincing deepfakes of real people, places, or events. The evaluated protections sit at the prompt and image layers.

On provenance, generated images carry C2PA metadata, and OpenAI layers an invisible watermark using Google DeepMind's SynthID across ChatGPT, Codex, and the API. OpenAI is also explicit that there is no single solution to provenance — these markers are layered signals, not absolute proof of origin.

For privacy, treat an uploaded reference image the way you would treat any document: consider whether the account is personal or business-managed, review who can see generated assets, remember that API use is separate from subscription use, keep API keys out of client-side code, and audit third-party integrations that receive image data. None of these specifics claim certifications or retention guarantees; they are practical defaults until you confirm the terms that apply to your workflow.

Who Should Use It?

Adopt ChatGPT Images 2.5 if you already generate images in ChatGPT or Codex — the upgrade arrives with the tool you have. Adopt it if you edit real photos and need the subject to survive a background swap, if you produce variations and campaign creative where consistency pays, or if you are a developer choosing an API model by job: Flare for speed and volume, Sunburst for edit control.

Hold off, or lower expectations, if your work is pixel-perfect masking in a professional raster editor, if you mainly want distinctive painterly fine-art output and should compare against dedicated art generators before committing, or if you expect identical results on identical prompts — no image model delivers that.

Frequently Asked Questions

What is ChatGPT Images 2.5? ChatGPT Images 2.5 is OpenAI's latest image generation and editing model, released on September 8, 2026 as the successor to Images 2.0. It powers image creation and editing inside ChatGPT, ChatGPT Work, and Codex, and ships separately as two API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.

What is new in ChatGPT Images 2.5? OpenAI says the model produces sharper detail with more natural lighting and textures, keeps reference-photo subjects more recognizable, and follows editing instructions more reliably across multiple turns. ChatGPT also adds Sketch, templates, on-image comments, and prompt sharing, with generation latency up to 50% lower than Images 2.0.

Is ChatGPT Images 2.5 better than Images 2.0? OpenAI positions ChatGPT Images 2.5 as the successor to Images 2.0 rather than a separate product: it appears in the same ChatGPT surface with no new subscription required. The gains it cites are sharper detail, more precise edits, better reference fidelity, steadier multi-turn consistency, and up to 50% lower latency.

What is the difference between GPT-Image-2.5 Flare and Sunburst? Flare is the default API model for most applications, pairing the quality and editing improvements with up to 50% lower latency than GPT-Image-2. Sunburst is built for premium creative workflows where tighter control across detailed edits matters more than speed, and it comes with longer generation times. Both bill at the same rates.

How much does GPT-Image-2.5 cost through the API? Per OpenAI's current pricing page, both GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst charge $5 per million text-input tokens, $8 per million image-input tokens, and $30 per million output tokens, with cheaper cached input. This is API pricing and is separate from any ChatGPT subscription.

Is ChatGPT Images 2.5 safe to use? OpenAI keeps the safety stack from Images 2.0 and adds graphical checks at the prompt and image layers, including a safety reasoning model that blocks violating inputs and outputs. Generated images carry C2PA metadata plus an invisible SynthID watermark, and OpenAI notes that no single provenance method is a complete solution.