What Is GPT Image 2.6? Markup Edits and Specs

Key takeaways

• GPT Image 2.6 is an OpenAI image model that has been named, but not released. Its only official mention is one line in OpenAI's API spec, added on October 28, 2026.
• That line describes a new edit field called markup: a transparent PNG you draw on to show the model where and how to change an image. The spec says it "Supports GPT Image 2.5 and 2.6."
• No model ID, price, resolution, reference limit or release date has been published. Any page showing a gpt-image-2.6 model ID in code is guessing.
• HiAPI has a GPT Image 2.6 page with a launch alert. GPT Image 2.5 is live today from $0.025 per image.

What is GPT Image 2.6?

GPT Image 2.6 is the next version number in OpenAI's GPT Image line, after GPT Image 2 (April 2026) and GPT Image 2.5 (September 8, 2026). It surfaced in a commit to OpenAI's public OpenAPI repository titled "Add 'markup' property to image editing endpoints" (commit aa880040). The new field's description reads:

A transparent PNG overlaid on the first image to guide the edit. Must match that image's dimensions after orientation and contain transparent pixels. Supports GPT Image 2.5 and 2.6. Cannot be combined with mask.

Those two sentences are everything OpenAI has said about 2.6. Across the whole spec, "2.6" appears only twice, both times in this description.

GPT Image 2.6 at a glance

How markup edits work

Image edits so far have had two ways to point at a region: describe it in the prompt, or send a mask, a PNG whose transparent area marks where the model may paint. Markup is a third way. You draw on a transparent layer the same size as the original, for example a circle around a logo, an arrow, or a short note, and the model reads those marks as edit instructions.

The spec sets four rules:

1. It's a PNG with transparency. Untouched areas must stay transparent.
2. It matches the first image's size after orientation. A phone photo stored sideways with an EXIF rotation counts at its displayed size.
3. It overlays the first input image. With several references, the marks apply to the first one.
4. It can't be combined with mask. Use one or the other in a request.

In JSON requests, markup can reference a file ID, an HTTPS URL or a base64 data URL; in multipart uploads, it's a file.

Build a markup PNG in Python

This script makes a valid overlay for any photo. It rotates the image by EXIF first, so the sizes match what the model will see, then checks both rules before saving.

from PIL import Image, ImageDraw, ImageOps

# Rotate by EXIF first, so the size matches what the model sees
original = ImageOps.exif_transpose(Image.open("original.jpg"))
w, h = original.size

# Fully transparent canvas, same size as the original
markup = Image.new("RGBA", (w, h), (0, 0, 0, 0))
draw = ImageDraw.Draw(markup)

# Circle the area to change, with a short label
box = (int(w * 0.30), int(h * 0.55), int(w * 0.70), int(h * 0.80))
draw.rectangle(box, outline=(255, 0, 0, 255), width=max(4, w // 150))
draw.text((box[0], box[1] - 24), "change badge to a star", fill=(255, 0, 0, 255))

markup.save("markup.png")

# Check the two rules from the spec
assert markup.size == original.size
assert markup.getextrema()[3][0] == 0, "needs transparent pixels"
print("ok", markup.size)

The model still needs a prompt. Write the change in words too ("Replace the lightning badge inside the red box with a white star; keep everything else") and treat the marks as the pointer.

What markup is likely to help with

These follow from how the field is defined, not from published 2.6 results:

• Product fixes: circle a label, a logo or a stray reflection and say what should replace it.
• Layout notes: mark where text or a product should move, instead of describing coordinates in words.
• Design review loops: a reviewer's red-pen marks on a mockup go straight into the edit request.
• Edits a mask can't express: a mask only says "paint here." Markup can carry arrows, labels and several notes at once.

How GPT Image 2.6 compares

Only one comparison can be made from official material: GPT Image 2.5 and 2.6 both accept markup. Whether 2.6 is faster, renders text better or supports more references hasn't been published. For what 2.5 already does, see GPT Image 2.6 vs GPT Image 2.5.

Availability: how to access GPT Image 2.6

OpenAI hasn't opened GPT Image 2.6 to the API. On HiAPI:

1. Open the GPT Image 2.6 model page and click Notify me at launch while signed in. You'll get an alert when it opens.
2. In the meantime, build on GPT Image 2.5 Flare or GPT Image 2.5 Sunburst, from $0.025 per 1K image. Request code is in the GPT Image 2.5 API guide.

HiAPI's GPT Image 2.5 requests don't include a markup field yet. New accounts get up to 2,000 free credits ($1).

What we don't know yet

• The model ID and whether there are variants, like 2.5's Flare and Sunburst.
• Price per image and whether it differs by resolution or quality.
• Maximum resolution, reference image limit and speed.
• Whether 2.6 is the model behind the Aurora or Shooting Star names seen in Arena. Nothing public links them; the sightings are covered in GPT Image 2.6 Leak.
• The release date. What to watch for is in GPT Image 2.6 Release Date.

FAQ

Is GPT Image 2.6 released?

No. It's named in OpenAI's API spec, but it isn't in OpenAI's model list and can't be called.

What is the GPT Image 2.6 model ID?

It hasn't been published. Code samples using gpt-image-2.6 are guesses and will fail.

Can I use markup today?

The spec lists GPT Image 2.5 as supported, so it's documented for 2.5 on OpenAI's own API. HiAPI's 2.5 requests don't include markup yet.

How much will GPT Image 2.6 cost?

No price has been published. GPT Image 2.5 on HiAPI costs $0.025 per 1K image, $0.04 at 2K and $0.06 at 4K, but that isn't a guide to 2.6 pricing.

What to watch next

The signals that 2.6 is close: a gpt-image-2.6 entry in OpenAI's model list, a changelog post, or the ID appearing in the edit endpoint's model enum. We'll update this page when any of them happens. To hear first, turn on the alert on the GPT Image 2.6 model page.