Équipe GPT Image 2.6The short answer: in GPT Image 2.6, a transparent background comes from the request, not the prompt. Set background to transparent and ask for PNG or WebP, and the file keeps a real alpha channel. Wording alone will not do it, and some renders paint a grey checkerboard into the pixels, which looks like transparency and is not. Check for clear pixels, view the file on light and dark, and note that 2.5 supports the same setting.
One field turns the background off
background is the switch, and it takes three values. transparent leaves the canvas empty, opaque fills it, and auto lets the model decide from your prompt. The field only works with a format that can store alpha, so the output format matters just as much.
OpenAI's docs state the pair directly: for a transparent background, set background to transparent and use output_format png or webp. Both 2.6 variants support it, gpt-image-2.6-flare and gpt-image-2.6-sunburst.
| Setting | Values | What it changes |
|---|---|---|
background |
auto, opaque, transparent |
Whether the canvas gets a fill |
output_format |
png, jpeg, webp (generations also take svg and gif) |
Whether alpha can survive |
quality |
auto, low, medium, high, xhigh, max |
Detail only |
complexity |
low, medium, high |
Planning, defaults to high |
Prompting still matters, but for the edges rather than the cutout. OpenAI's logo example asks for "a fully transparent background" and then rules out "solid backdrop, scenery, checkerboard, or watermark". Those words ask for clean edges and no backdrop. They do not replace the field.
Pick a format that stores alpha
Choose a format before you generate, because one wrong choice throws the alpha away. PNG and WebP keep it. JPEG has no alpha channel at all, so a clean cutout saved as JPEG arrives with a solid background.
Format choice also has endpoint limits. SVG and GIF output exist only on the generations endpoint; the image edit endpoint supports neither.
output_format |
Keeps alpha | Works on edits |
|---|---|---|
png |
Yes | Yes |
webp |
Yes | Yes |
jpeg |
No | Yes |
For a still cutout, PNG is the safe default and WebP is the smaller one. output_compression applies to JPEG and WebP only, so PNG output ignores it.
Check the file, not the request
A correct request can still hand you a file that only looks transparent, so test the pixels. Three checks take about a minute.
Look for a grid. Many viewers draw their own checkerboard behind clear pixels, so the grid alone proves nothing. Paste the file onto a solid colour instead. Tiles that move with the image are painted pixels.
Ask the file for clear pixels. Any editor with an alpha view shows them. On macOS, one line also works:
sips -g hasAlpha out.png
It prints hasAlpha: yes or hasAlpha: no. If the answer is no, no later step can add the missing pixels back.
Place it on white, then on black. Soft edges and light shadows look clean on white and glow on dark. Hair, glass and thin outlines show the problem first, so check those on both.
A 1024×1024 WebP with a real alpha channel; the background around the sticker is clear, and only its soft shadow is partly transparent.
When the background comes back
Failures come from four places: a prompt that asks for a scene, a format that drops alpha, an edit that restores the old backdrop, or a model that draws the checkerboard itself.
| What you see | Why | Fix |
|---|---|---|
| Grey tiles inside the image | The model drew them. A painted checkerboard is not transparency | Rerun and describe only the subject |
| A solid colour behind the subject | background stayed on auto, or the prompt named a backdrop |
Set transparent and remove scene words |
| A white box around a cutout | The output was JPEG | Switch to PNG or WebP |
| The edit brings the backdrop back | The input image had one | Ask for a transparent background again |
Edits are the weak spot, because the model starts from a picture that already has a background. OpenAI's own product cutout example repeats the transparency request on later edits for that reason. Markup overlays have a related rule: the PNG you upload must be transparent and match the first input image's pixel size.
This is not new in 2.6
Transparency shipped before 2.6, so it is not a reason to upgrade. OpenAI's 2.5 prompting guide documents the same background field and the same alpha check, including the warning to inspect hair, glass and edges.
What 2.6 adds is elsewhere: a complexity dial, SVG output and animated GIF output. Token rates are unchanged at $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens. GPT Image 2.5 also has a Batch rate of $15 per million image output tokens; 2.6 does not appear in that Batch table, and its model pages list no Batch price. Equal rates still do not mean an equal cost per image, and the 2.6 vs 2.5 comparison covers that split in full.
Run the prompt in a browser
The browser generator here runs both variants, with no OpenAI account or API key. It is an easy place to settle a prompt before you spend API calls. Pick Flare for quick passes or Sunburst when detail has to hold, at 1K, 2K or 4K. Quality is low, medium or high, and high adds 50% to the credit price.
A still costs 20, 30 or 60 credits on Flare, or 26, 39 or 78 on Sunburst. An animated GIF is 40 credits on Flare and 52 on Sunburst, and you can attach up to 16 reference images. Failed runs refund their credits.
Use the browser to settle wording, framing and size. Transparency is decided by the request fields above, so set background and output_format on whichever endpoint produces the final file. The API reference lists both model IDs and the token rates, and the pricing page covers credit packs.
FAQ
Does GPT Image 2.6 support transparent backgrounds?
Yes. Both gpt-image-2.6-flare and gpt-image-2.6-sunburst accept background: "transparent". The output has to be PNG or WebP, because JPEG cannot store alpha.
Is transparency new in 2.6?
No. GPT Image 2.5 supports the same background field and the same alpha check. 2.6 adds complexity control, SVG output and animated GIF output, not transparency.
Why is there a checkerboard in my PNG?
The model drew those tiles into the picture. They are ordinary grey and white pixels, so no setting removes them afterwards. Generate again and describe the subject without naming a backdrop or a pattern.
Which format should I use for a cutout?
PNG for a lossless file, WebP for a smaller one. Both keep alpha. SVG output exists only on the generations endpoint, so it is not an option for edits.
Can I edit an image and keep the background transparent?
Sometimes, and you have to ask on every edit. If the input image has a background, say again that you want it gone, then check the new file for clear pixels.
What is the model ID for GPT Image 2.6?
Use gpt-image-2.6-flare or gpt-image-2.6-sunburst. There is no bare gpt-image-2.6 model ID, so a request has to name one of the two variants.
Sources
- Image prompting guide — the 2.6 parameter list, both model descriptions, the alpha check and the "no checkerboard" prompt line.
- Image generation guide — the transparent plus PNG or WebP rule, format and endpoint limits, and the complexity default.
- Create image API reference — transparent background support by model and the PNG or WebP requirement.
- API pricing — standard rates for 2.6 and 2.5, and the Batch rates that list 2.5 only.
- Changelog — the November 4, 2026 release note.
- gpt-image-2.6-sunburst and gpt-image-2.6-flare — model pages and token prices.
- OpenAI Developer Community thread — user reports of a checkerboard baked into the image.
Revision log
- November 5, 2026 — Page published.