Release Watch
EvoLink TeamProduct Team7 min read
GPT Image 2.6 Leak: OpenAI’s Docs and Anonymous Model Clues
What OpenAI’s markup documentation says, and what Aurora and Shooting Star sightings do—and do not—tell us about GPT Image 2.6.
EvoLink TeamProduct TeamOn this page
GPT Image 2.6 has appeared by name in OpenAI's image-editing documentation. A new
markup parameter describes support for both GPT Image 2.5 and 2.6. That is a concrete clue about the name and an editing interface, but it is not a public model launch: as of October 29, 2026, OpenAI's model catalog and API changelog do not list a 2.6 release or public model ID. Image edit reference, model catalog, API changelogMeanwhile, people have shared images from experimental models called Aurora and Shooting Star. Some look promising, but OpenAI has not identified either as GPT Image 2.6. Their samples are useful for seeing what testers are trying; they cannot establish 2.6's performance, price or launch date.
What is confirmed about GPT Image 2.6?
On October 28 at 20:56 UTC, an OpenAI OpenAPI repository update added
markup to image-editing specifications. The same description appears in the online API reference and names both 2.5 and 2.6 as supported versions.| Question | What is known as of October 29 |
|---|---|
| Has the name appeared in official documentation? | Yes, in the image-editing markup description. |
| Is there a public API model ID? | No 2.6 ID is listed in the checked public model catalog or image-edit reference. |
| Is there an official launch date or price? | Neither was found in the checked official sources. |
| Is Aurora or Shooting Star confirmed as 2.6? | No official mapping has been published. |
| Is it available on EvoLink? | The EvoLink 2.6 page is a coming-soon page, not a live API listing. |
The absence of an entry in the sources checked does not describe access held by every private tester. It does mean a guessed model ID or a third-party claim should not be treated as an official integration contract.
What does the new markup parameter do?
OpenAI describes
markup as a transparent PNG overlaid on the first input image to guide an edit. The overlay must contain transparent pixels and match the image's dimensions after orientation. It can be supplied by file ID, HTTPS URL or base64 data URL, and it cannot be combined with mask in the same request. OpenAI image edit reference
Screenshot shared by Token Gremlin; verify the parameter in the OpenAI API reference.
For example, a designer could mark a product photograph to indicate the area that needs attention, then send the original image and overlay with an editing instruction. That is a possible workflow based on the input format, not a demonstrated 2.6 result. The reference does not specify how every arrow, color or handwritten note will be interpreted, and it does not guarantee that pixels outside the marked area remain unchanged.
The important detail is that the description names both 2.5 and 2.6. Markup is not evidence of a 2.6-exclusive feature, and a description in a specification does not tell us when an account will be able to call 2.6.
Aurora and Shooting Star: what the samples show
Aurora appeared in a model list and in Arena images
CellCog reported that an authenticated account saw
aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha in an OpenAI model list on October 18. Its author said they inspected the list rather than successfully calling those models. A listing by itself says little about access, quality or cost.LuminaBench later shared Arena captures labelled Aurora Red and Green. Both depict a tabby cat: the red example is a tighter portrait, while the green example shows more of the room and body. Two selected images cannot establish which variant is better, faster or closer to a future release. The names in the samples also do not prove that Aurora is GPT Image 2.6.
Shooting Star samples are not a confirmed 2.6 benchmark
Community testers also shared outputs labelled
shooting-star-red-alpha and shooting-star-green-alpha. Some attributed them to OpenAI based on the images or third-party analysis, but the discovery material did not identify a confirmed provider. No public source maps those labels to 2.6. Community discoveryOne intersection test asked for seven cars, three bicycles, five pedestrians and two buses, plus directions, shadows and an occlusion. The shared results illustrate why a realistic-looking scene and a correct scene are different things: the visible cars did not fully match the requested count. These are individual outputs, not a controlled success rate or a measurement of GPT Image 2.6.
A style-grid post compares two images. An index of the author’s posts describes a comparison labelled GPT Image 2.5 Sunburst and Shooting Star Green, but the original post could not be accessed to verify that the index refers to these exact images. The model labels therefore remain unconfirmed for this pair. The available materials do not establish matching style lists, settings, attempt counts or image selection, so this is not a reliable head-to-head test.
GPT Image 2.6 vs 2.5: what can actually be compared?
GPT Image 2.5 already has two public API models, Flare and Sunburst. OpenAI positions Sunburst for editing precision and Flare for faster everyday image generation. Both support
xhigh and max quality settings; the current image-edit reference accepts up to 16 input images and describes transparent output for 2.5. OpenAI image generation guide, image edit referenceThose are 2.5 capabilities, not confirmed 2.6 specifications. Likewise, the
markup description includes 2.5. Until OpenAI publishes a 2.6 model card and testable access, there is no verified basis for saying that 2.6 adds 4K output, supports more reference images, fixes text rendering, or is faster than either 2.5 model.For anyone evaluating an image model, the useful comparison is a set of repeatable tasks: preserve a product's identity through several edits, render the same short and dense text, count and place multiple objects, and measure total time and cost including retries. A single impressive picture cannot answer all of those questions.
Availability and pricing
OpenAI had not published a 2.6 price or public model ID in the sources checked on October 29. For current access information and updates, see the GPT Image 2.6 model page.
Frequently asked questions
When will GPT Image 2.6 be released?
No official date was found in the sources checked on October 29. Posts predicting an imminent launch are speculation, not a schedule from OpenAI.
Can I call GPT Image 2.6 in the API now?
The public sources checked on October 29 did not list a 2.6 model ID. See the model page for access updates.
Is markup the same as a mask?
They are separate request inputs. The reference describes markup as a transparent PNG overlay to guide an edit and says it cannot be combined with
mask. It does not fully explain how every type of visual annotation is interpreted.Is Shooting Star or Aurora GPT Image 2.6?
There is no official confirmation linking either experimental name to 2.6. Public samples should retain the labels under which they were shared.
Is GPT Image 2.6 better or cheaper than 2.5?
There is not enough verified public information to answer either question. Performance, price and release status need to be established separately when OpenAI publishes them.
