Your credits refresh every month for as long as ListingRVA AI exists. No renewals, no annual increases, nothing to cancel.
Available on select plans. No renewals, ever.
November 9, 2026 · 11 min read · by Aashirvad Kumar
You ask an AI tool to put your serum bottle on a clean white background, and the edit comes back with one letter missing from the label, a slightly taller cap and an amber that is a shade too orange. Nobody spots it until the "not as pictured" returns and one-star reviews start, and a tool still running on a retired OpenAI model can stop generating in the same week you need new photos. GPT Image 2.6 is OpenAI's new image model family, announced on 4 November 2026 in two API versions called Sunburst and Flare, and it does not guarantee that your label, logo or colour survive an edit, so the safe setup is Sunburst for edits of your real photo, a deliberate complexity setting and a side-by-side check of every output.
This guide is for sellers who edit their own product photos with AI, either directly through OpenAI's API or through a tool built on it. It covers what changed, which version suits product work, the default setting that can introduce unwanted changes, what the new models cost, what happens on 1 December and a check to run before every upload. It does not cover the new ChatGPT app features, and it does not rank 2.6 against other companies' models, because we have not seen a published benchmark to rank it on. Colour drift is only one reason buyers send things back, as the breakdown of not as pictured returns explains.
Key takeaways
GPT Image 2.6 is a pair of OpenAI image models, Sunburst and Flare, that both generate new images and edit existing ones through the API. OpenAI's launch post on its developer forum, published on 4 November 2026, introduced them as coming to the API "with 50% faster image generation", and described each one in a line: "Flare generates images in as little as six seconds" and "Sunburst brings higher-quality generation and greater consistency across edits."
For a seller, the phrase that matters is "consistency across edits". A typical product photo goes through several passes: remove the background, add a soft shadow, extend the canvas, then change the scene for a lifestyle shot. Each pass is a chance for the model to redraw part of the product instead of keeping it, and the small things are the easiest to lose: the letters on a label, the curve of a logo, the exact shade of a cap. Greater consistency means fewer of those changes between passes. It does not mean zero, and OpenAI has not published a number for how much more consistent the new models are.
OpenAI's Sunburst model page describes it as "Image generation and editing for detailed visual work" and lists support for the v1/images/generations and v1/images/edits endpoints, with Chat Completions, Responses and Batch not supported. That last point matters to tools that process a catalogue in bulk, because a job built on Batch cannot simply switch to 2.6. Whether the edit is a background swap or a whole new scene, the starting point is a good source photo, which the AI product photo workflow guide walks through step by step.
Use Sunburst when you are editing your real product photo and Flare when you want fast drafts of an idea, because OpenAI's image generation guide describes Sunburst as "for workflows where editing precision matters most" and Flare as "fast, high-quality everyday image generation". Both carry the same token prices, so the choice is about the job, not the bill.
| What you need to know | Sunburst | Flare |
|---|---|---|
| OpenAI's description | For workflows where editing precision matters most | Fast, high-quality everyday image generation |
| Seller use | Editing your real product photo: main image, white background, A+ module images | Scene ideas, moodboards and drafts you will not upload as the product |
| Speed claim | Not stated separately | Images in as little as six seconds |
| Token prices (per 1M) | Text input $5, image input $8, image output $30 | Same as Sunburst |
| Endpoints | v1/images/generations and v1/images/edits | Same as Sunburst |
| Complexity | Low, medium or high; high if omitted | Low, medium or high; high if omitted |
| Quality options | Low, medium, high, xhigh, max | Low, medium, high, xhigh, max |
In practice, anything a buyer will compare with the box on their doorstep belongs on Sunburst: the main image, the white background shot, close-ups of the label and A+ module images that show the real pack. Flare suits work where speed matters more than precision, such as testing five kitchen scenes before you pick one, or drafting a seasonal banner idea. If a Flare draft wins, recreate it as an edit of your real photo rather than uploading a generated product. Any main image still has to pass the Amazon main image rules.
GPT Image 2.6 does not guarantee that your label, logo or colour survive an edit: OpenAI promises greater consistency across edits with Sunburst, not an unchanged product, so every output still needs a side-by-side check. The biggest lever you control is the complexity setting, and if you edit through a tool rather than the API, you may never see it unless you ask.
According to OpenAI's image generation guide, complexity accepts low, medium or high, and if it is omitted both models use high. The same guide warns that "Higher complexity can increase latency and introduce unwanted changes", and says adjusting complexity has no separate fee. For a product edit, unwanted changes are exactly the failure you are trying to avoid. A simple background swap or shadow does not need the model to rethink the whole image, so a lower complexity is worth testing against the default on five of your own photos before you change anything in bulk.
The default is high: if your tool does not set complexity, both 2.6 models run at high, the setting OpenAI says can introduce unwanted changes. Ask your tool which value it sends for edits.
Quality is a separate setting from complexity. Sunburst and Flare both offer low, medium, high, xhigh and max, and a higher quality value is not the same as a more faithful edit. Two other gaps are worth knowing. We have not seen a published benchmark or quality score for 2.6, so treat any claim that it is better for product photos as untested until you try it on your own products. OpenAI's documentation also does not say whether 2.6 supports the input_fidelity parameter, so do not assume a tool can switch it on. Background colour drifts as well as label colour, and a white that turns light grey is the off-white background problem.
Some edits carry more risk than others. Fine print, ingredient panels, barcodes, embossed logos, transparent bottles and metallic finishes are where a redraw is hardest to spot at thumbnail size and easiest for a buyer to notice in hand. Photograph those products with the label square to the camera and well lit, because a sharp, readable original gives any model less to guess.
OpenAI does not publish a flat per-image price for GPT Image 2.6, so the cost of one product photo depends on how many output tokens it uses. The Sunburst model page lists $5 per 1M text input tokens (cached $1.25), $8 per 1M image input tokens (cached $2) and $30 per 1M image output tokens, and Flare carries the same prices. OpenAI's guide says to read the usage field on each response to see what a job actually cost.
That makes your own test the only reliable number. Run ten real edits at the settings you plan to use, read the usage field and divide. One early report suggests the count can differ from 2.5: one developer on the OpenAI forum reported that a medium-quality 2.6 request billed 1136 output tokens against 344 for 2.5 on the same prompt. That is one person's test, not an official figure, but it is a reason to measure before you move a whole catalogue. If you pay a tool in credits rather than paying OpenAI directly, the token price is the tool's cost, so the question for you is whether the tool changed its credit price per image.
Your tool will stop generating on 1 December 2026 only if it still calls gpt-image-1.5, gpt-image-1-mini or chatgpt-image-latest and has not moved to a replacement. OpenAI's deprecations page says: "On June 2, 2026, we notified developers using older GPT Image models of their deprecation and removal from the API on December 1, 2026." It names gpt-image-2.6-sunburst or gpt-image-2.6-flare as the recommended replacement, and notes that gpt-image-1 was shut down on October 23, 2026.
Photos you already downloaded and uploaded are ordinary files, so the removal affects new edits, not your live listings. Most tools do not say which model they run on, so ask four questions before launch week:
The third question matters most for product accuracy. An edit starts from your pixels and changes part of the picture, while a fresh generation draws the product again from a description or a reference, which gives the model far more room to change the label, logo and shape. A tool that cannot tell you which one it does is a tool whose output you need to check twice. Amazon's own use of AI-made images is a separate topic, covered in the post on AI images in Amazon search.
On 8 November 2026, ListingRVA AI switched its OpenAI image model to gpt-image-2.6-sunburst, chosen over Flare because its images are edits of the seller's own product photo and Sunburst is OpenAI's editing-precision model at the same token price. If 2.6 fails or times out, the same job retries once on gpt-image-2.5-sunburst, and requests always ask for quality high.
Two tradeoffs come with that setup. First, ListingRVA AI picks the model per image style automatically from several AI models and switches to a backup if one fails, so not every image is made by GPT Image 2.6, and images from different models can look different. Second, we have not measured whether 2.6 improves label or colour accuracy, so this post makes no such claim. What did not change: credit prices per image stayed the same, and because ListingRVA AI does not use gpt-image-1.5, the 1 December shutdown does not affect it. None of this removes the need to check every image yourself, and the styles and export sizes it covers are on the ListingRVA AI features page.
Check every AI edit against the original photo at full size before upload, reading the label letter by letter and comparing the logo, colour, shape and every part of the product. The check costs far less than one return. Open the original and the edit side by side at the same zoom and work through this list:
Fix a failed edit by redoing it, not by patching it, and lower the complexity or narrow the instruction before you try again. If the only job is a cleaner background, a cut-out step is a different task from a generative edit, because it is meant to separate the product rather than redraw it, which is what a product image background remover is for.
Quick test: flip between the original and the edit in the same window. Anything that jumps when you flip has changed, and anything on the product that changed should not ship.
Keep the original file next to every edit in your folder, named so the pair is obvious, because the original is your proof of what the product really looks like. If an edit adds a person to a lifestyle shot, a separate Amazon rule applies on top of these checks, set out in the guide to the Amazon synthetic performer rule.
Create white background, lifestyle, infographic and model shots, A+ pages and listing copy for Amazon, Shopify, Etsy, Flipkart and Meesho. Start free with 50 credits, no credit card.
Start free →Not guaranteed. OpenAI says Sunburst brings greater consistency across edits, which is a claim about fewer changes, not a promise of none. Compare every edit against your original photo at full size, reading the label letter by letter and checking the logo, colour and parts, before you upload it.
Use Sunburst when you edit your real product photo, because OpenAI's image generation guide positions it for workflows where editing precision matters most. Use Flare for fast drafts and scene ideas that will not be uploaded as the product itself. Both cost the same per token.
Complexity accepts low, medium or high, and both 2.6 models use high if your tool does not set it. OpenAI says higher complexity can increase latency and introduce unwanted changes, and changing it has no separate fee. For simple edits such as a background swap, test a lower value against the default.
OpenAI does not publish a flat per-image price. Sunburst and Flare are billed at $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens, so the cost of each image depends on its output tokens. Read the usage field on a few real jobs to see your number.
Only if it still calls gpt-image-1.5, gpt-image-1-mini or chatgpt-image-latest. OpenAI removes those from its API on 1 December 2026 and recommends gpt-image-2.6-sunburst or gpt-image-2.6-flare instead. Ask your tool which model it uses and what its backup is.
No. ListingRVA AI picks the model for each image style automatically from several AI models and switches to a backup if one fails. Its OpenAI image model is now gpt-image-2.6-sunburst, with a retry on gpt-image-2.5-sunburst, but other images come from other models. Check every image before you upload it.
ListingRVA AI now works inside Claude. Connect your workspace and ask Claude to generate listing copy and publish product images to your store, without leaving your editor. See the Claude connector →
Comments
No comments yet, be the first.
Leave a comment