OpenAI's Two New Image Models Edit One Part of Your Product Photo Without a Reshoot

Two New Image Models From OpenAI and What They Change for Product Photos

On September 8, 2026, OpenAI added a new generation of image models under the name GPT Image 2.5, and it shipped as two models rather than one: gpt-image-2.5-sunburst, the more capable of the pair, built for workflows where editing precision matters most, and gpt-image-2.5-flare, the faster one for high-quality everyday image generation. The practical takeaway for a business owner: you can now edit a specific region inside a product photo you already own, instead of booking a studio reshoot or regenerating the whole image and hoping it comes out similar.

What OpenAI Actually Announced

According to the model pages in OpenAI's official documentation, both models accept text and image inputs and produce image outputs only. Each carries a pinned snapshot dated to the launch day itself: gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08. That detail matters operationally, because pinning a snapshot means your output does not shift underneath you the day the provider updates the model.

Both models support six quality settings: low, medium, high, xhigh, max, and auto. And both work through two paths: the Image API with its generation and edit endpoints, or as a tool inside the Responses API when you want a multi-turn experience where a user refines the image by talking to it. The first path is enough if you need one image from one prompt. The second is what you need if you are building a screen where a marketing employee edits an image step by step.

Sunburst or Flare? The Rule Is Simple

Do not choose by name, choose by the nature of the work. If the job demands editing precision — changing the color of a bottle, removing an element from a background, swapping text on a package, keeping the product shape exactly as it is — that is Sunburst territory, the model OpenAI describes as its most capable for image generation and editing. If the job is about volume and speed — dozens of campaign backgrounds, illustrations for articles, quick visual drafts before a direction is approved — Flare is the logical pick.

In real use most teams need both: Flare during exploration where quantity is the point and mistakes are cheap, then Sunburst for the final pass where the image lands on a product page a paying customer will see.

Partial Editing Is the Real Change, Not Generation

Text-to-image generation is neither new nor rare anymore. What deserves attention here is inpainting support — redrawing a defined region inside an existing image while the rest of the image stays untouched. That turns the tool from an image generator into an image editor, and the difference between the two is the difference between something that entertains the marketing team and something that enters an actual production line.

A concrete example: a store with three hundred SKUs all shot on a white background decides to move to a light grey visual identity. The old route is scheduling a new shoot, moving inventory, paying a photographer, two weeks. The new route is running the images through the API with one instruction, adding a human review of the results, one working day. Same photo, same product, only the surroundings changed.

Cost Is Counted in Tokens, Not in Images

This point confuses many business owners because it does not resemble design-service pricing. According to OpenAI's pricing for both models, the rate is identical across the pair: text input at five dollars per million tokens and one dollar twenty-five cached, image input at eight dollars per million tokens and two dollars cached, and image output at thirty dollars per million tokens. Text output is not billed at all, because these models output images rather than text. OpenAI also notes that the token rates match the previous GPT Image 2 generation.

The practical advice: do not estimate on paper. Run a trial batch of a hundred real images from your own catalog with the exact settings you intend to use, read the invoice, then multiply. An estimate built on a hundred real images is more honest than any calculator, especially since the quality level you pick directly changes consumption.

Five Uses That Pay Off in the First Week

• Catalog consistency: uniform backgrounds, lighting, and angles across hundreds of items shot at different times and under different conditions.
• Campaign variants: a summer version, a Ramadan version, and a White Friday version of the same photo without a shoot for every occasion.
• Packaging preview before print: test Arabic copy or a new logo on the package visually before committing to a print run.
• Restoring old images: remove distracting elements, reflections, or outdated marks from archive photos instead of discarding them.
• Content imagery: illustrations for your blog and newsletter instead of the same stock photos your competitors also use.

Three Notes Before You Build On It

First, organization verification. OpenAI's documentation states you may need to complete API organization verification from your developer console before using GPT Image models. Put that step into the timeline from the start so it does not ambush you on launch day.

Second, human review is not optional. Any product image that will be published needs a person signing off on it. The model does not know that the color it adjusted is the one registered in your brand identity, and it does not know the edit made the product look bigger than it is. Showing a product with an image that does not match reality is a commercial and regulatory problem before it is a technical one.

Third, do not upload photos of people. Customer or employee photos are not raw material for experiments. Make the rule explicit for your team: products, places, and marketing material yes; people no, unless you have documented explicit consent under a clear personal data policy.

How This Enters Your System Instead of Staying a Side Tool

The difference between a company that benefits from these models and one that merely talks about them is where the models are used. If they live in a browser tab an employee opens to download a file by hand, the gain is small and disappears with the first vacation. The value appears when you wire the API into your system: a new item enters the catalog, its visual variants are generated automatically at the sizes each channel needs, they enter a review queue, and they publish with one approval click. At that point you are not buying images, you are building a visual production line that runs without daily supervision.

At Origami we build exactly this kind of wiring into the systems we develop for clients: an API connected to the product database, a review queue with clear permissions, and a log showing who approved which image and when. The technology is available today, and the deciding factor is the decision itself: do you treat it as a new toy, or as a step in an operational process with an owner and a success metric.

Sources

• OpenAI official documentation — GPT-Image-2.5 Sunburst model page: developers.openai.com
• OpenAI official documentation — GPT-Image-2.5 Flare model page: developers.openai.com
• OpenAI image generation guide: developers.openai.com