GPT Image 2.6 vs GPT Image 2.5: What Changes?

GPT Image 2.6 is not out. As of November 1, 2026, OpenAI lists no gpt-image-2.6 model ID, price or changelog entry. The only official mention is in its API spec, which described a new edit field, markup, three times as "Supports GPT Image 2.5 and 2.6." On October 31, OpenAI cut 2.6 from two of the three; one still names it.

What you can use today is GPT Image 2.5, in two versions: Sunburst and Flare.

Short answer

• 2.6 is not released: no model ID, price or date.
• The only change OpenAI ties to 2.6 is markup, a new edit field that the spec also lists for 2.5. OpenAI's image guide doesn't mention it yet.
• In our 10 edits, Sunburst and Flare tied on recoloring and chalk text, and only Sunburst followed notes drawn on the photo, in 1 of 2 runs.
• Use Sunburst for edits and Flare for everyday images at the same token price, and write edit instructions in the prompt, not only on the image.

Is GPT Image 2.6 out?

No. We checked OpenAI's official pages on November 1, 2026.

• The model list shows two image models, Sunburst and Flare, both 2.5, and the full catalog links eight image pages, none of them 2.6.
• The 2.6 model page address returns "Page not found".
• The last image-model entry in the changelog is the 2.5 launch on September 8, 2026, and nothing about images was added on October 28.
• The pricing page has no 2.6 row.
• The API spec lists 10 image model IDs, from gpt-image-1.5 to chatgpt-image-latest, and none is gpt-image-2.6.

Where does "GPT Image 2.6" come from?

From one sentence in OpenAI's API spec, added on October 28 and partly removed on October 31.

The October 28 commit added three descriptions of markup, each ending "Supports GPT Image 2.5 and 2.6. Cannot be combined with mask." The October 31 commit changed two to "Supports GPT Image 2.5." and left the third. So the two request descriptions say 2.5 only, while the reusable markup schema in the same file still names 2.6. OpenAI has not said why.

The Python SDK still carries the older wording: its commit landed on October 29, v3.42.0 shipped on October 31 at 00:26 UTC, and the spec changed about 17 hours later.

On October 29, 2026, WinCentral wrote, citing no source, that 2.6 will fix text, hands and object placement and that a launch looks imminent; OpenAI has not confirmed it, and its spec says nothing about image quality.

What changes in 2.6, according to OpenAI?

Only one thing: a new edit field called markup, which the spec lists for 2.5 as well.

• A transparent PNG laid over the first image to guide the edit.
• It must match that image's dimensions after orientation and contain transparent pixels.
• It cannot be combined with mask.
• It sits on POST /v1/images/edits only, as a multipart file or a JSON reference: file ID, HTTPS URL or base64 data URL.
• The spec states no size cap, while a mask must be under 4 MB.

The field appears zero times in OpenAI's image generation guide, changelog, the two 2.5 model pages and the pricing page, all checked November 1, 2026. OpenAI gives no example of a markup image and doesn't say how the model reads one, and the release date, price, speed, sizes and quality levels of 2.6 are all unpublished.

2.5's own launch is easier to pin down: the September 8, 2026 changelog entry added the xhigh and max quality levels and kept GPT Image 2's token rates. For sizes and quality levels side by side, see our GPT Image 2.5 vs GPT Image 2 comparison.

How do GPT Image 2.5 Sunburst and Flare edit today?

Written instructions were followed in 6 of 6 edits, and notes drawn on the photo in 1 of 4.

How we tested

• Models: gpt-image-2.5-flare and gpt-image-2.5-sunburst, on their 2026-09-08 default snapshot, tested November 1, 2026.
• Settings: the edit operation, 1536x1024, quality medium, one image per request, the source as an inline Base64 JPEG at most 1024 px on the long side, since OmniaKey caps an edit request at 1 MiB.
• The source photo came from Nano Banana 2.1, so neither model edits its own picture. For test 2 we drew red notes on a copy: a circle around the cookie jar with REMOVE, and ADD A RED APPLE with an arrow to the empty counter.
• Runs: one per model in test 1 and two in tests 2 and 3, so 10 finished edits. One Flare request timed out at 186.7 s and was resent.
• Time runs from sending the request to the status check that saw the image, polled every 4 seconds, so each figure can be up to about 4 s high.
• We put every result next to the source at 2× zoom and checked pixel differences where no prompt touches.
• OmniaKey sells API access to these models, so we ran the tests through it. Every price is OpenAI's own rate, not ours. OmniaKey has no markup field yet, so markup itself is untested.
• Sources: OpenAI's Sunburst and Flare model pages, the changelog, the image guide and the pricing page, all checked November 1, 2026.

Ten edits is a small sample: read each result as one example.

Test 1: Recolor one object (3:2)

This test checks one clear color change and whether the rest of the frame survives.

Prompt:

Change only the white mug on the left into a bright yellow ceramic mug. Keep everything else in the photo exactly the same: the cookie jar, the fern, the chalkboard and its text, the counter, the wall and the light.

• GPT Image 2.5 Flare (20.7 s): a bright lemon-yellow mug, same shape, handle and position, but glossy where the original was matte. The jar, fern, chalk text, counter and wall stayed in place, while the chocolate chips on the cookies moved.
• GPT Image 2.5 Sunburst (20.2 s): a deeper golden-yellow mug, also glossy, with a slightly thicker handle. The rest stayed in place, and the chips moved here too.

Verdict: Tie. Flare's yellow is closer to the bright yellow we asked for and Sunburst's is more golden, and both redrew small details, so zoom in after an edit.

Test 2: Follow notes drawn on the photo (3:2)

This test checks whether the models read notes drawn on the image and clean up the marks.

Prompt:

This photo has red hand-drawn notes on it. Do what the notes ask, then erase every red mark and every red word so the result looks like an untouched photo. Keep everything else in the photo exactly the same.

• GPT Image 2.5 Flare (22.2 s and 22.0 s): both runs erased the red circle, words and arrow completely, then stopped. The jar stayed and no apple appeared.
• GPT Image 2.5 Sunburst (26.9 s and 22.2 s): run 1 removed the jar, filled the gap with clean wall and counter, and put a red apple with a stem and a natural shadow where the arrow pointed. It also redrew the fern denser and brought out the wall's plaster texture: on the wall, which no prompt touched, its average pixel difference from the source is 8.2 on a 0–255 scale, against 3.6–3.8 in the other three test-2 edits. Run 2 only erased the marks, like Flare.

Verdict: Sunburst, narrowly: it did the whole job in 1 of 2 runs and Flare in 0 of 2, though both erased every red mark. Write the change in the prompt too.

Test 3: Change chalk text (3:2)

This test checks two lines of chalk handwriting: same words, same place, same size.

Prompt:

Change the chalkboard text from FRESH BREAD to HOT SOUP, written in the same white chalk handwriting, at the same size and position, on two lines. Keep everything else in the photo exactly the same.

• GPT Image 2.5 Flare (22.3 s and 22.9 s): both runs wrote HOT and SOUP on two lines in white chalk at the original size and position, spelled right, with nothing extra. The letters are a little neater than the original handwriting, and run 2 left a faint erasure mark right of the P, where BREAD's D had been.
• GPT Image 2.5 Sunburst (32.0 s and 22.2 s): the same two lines, spelled right, at the same size and position, with letters a little neater too and no erasure marks.

Verdict: Tie, 4 of 4: both swap two lines of chalk text cleanly.

Results, price and speed

Test 1 was a tie on shades of yellow, Sunburst took test 2 by one run, and test 3 was a tie at 4 of 4.

OpenAI's pricing page, checked November 1, 2026, gives both the same rates.

Batch halves the image rates, to $4, $15 and $1 per 1M tokens. The rates equal GPT Image 2's, while GPT Image 1.5's image output costs $32 per 1M tokens.

The cached input rate applies only inside the Responses image tool, and direct /v1/images/edits calls pay the full $8 per 1M image input tokens. OpenAI publishes no per-image price for 2.5, and its model pages say the GPT Image 2 calculator doesn't estimate 2.5.

Speed matched: both medians were 22.2 s, Flare between 20.7 s and 22.9 s and Sunburst between 20.2 s and 32.0 s. OpenAI calls Flare its fastest image model, but in our edits the two were level. One report shows a big gain over GPT Image 2: on Hacker News on September 8, 2026, jjcm wrote that about 50k GPT Image 2 images had averaged "around 104s", while with 2.5 "images are coming in at around 35-40s."

Which should you use while you wait for 2.6?

Use Sunburst for edits and Flare for everyday images, at the same token price.

GPT Image 2.6 has no model ID yet, so there is nothing to switch to, and the one change OpenAI's spec names, markup, is a field on the same /v1/images/edits endpoint.

For a same-prompt test against Google's model, see how GPT Image 2.5 compares with Nano Banana 2.1.

How to call GPT Image 2.5 on OmniaKey

Both 2.5 models are on OmniaKey for generating and editing.

• quality takes low, medium, high or auto, while OpenAI's xhigh and max are left out.
• An edit takes an inline Base64 image, and the whole request is capped at 1 MiB.
• Several reference images go in images.
• There is no mask or markup field yet.

BASE64=$(base64 -i photo.jpg | tr -d '\n')

curl --fail-with-body https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer your-omniakey-api-key" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: mug-recolor-001" \
  -d '{
    "model": "gpt-image-2.5-sunburst",
    "type": "image",
    "operation": "edit",
    "input": {
      "prompt": "Change only the white mug into a bright yellow ceramic mug. Keep everything else the same.",
      "image": {"data": "'"$BASE64"'", "mime_type": "image/jpeg"}
    },
    "parameters": {"size": "1536x1024", "quality": "medium"}
  }'

1. The reply is 202 with a task ID.
2. Poll GET /v1/media/tasks/{id} until the status is succeeded.
3. Download the image from the signed URL in the task's assets.

Swap the model to gpt-image-2.5-flare for Flare. Current rates are on the Sunburst and Flare model pages, every parameter is in the image API docs, and you can try both in the AI Image Generator.

Frequently asked questions

Is GPT Image 2.6 released?

No. The model list (checked November 1, 2026) shows two image models, Sunburst and Flare, from September 8, 2026. A 2.6 page returns "Page not found", and neither the changelog nor the pricing page lists a 2.6 entry.

What is markup in the OpenAI image edit API?

An optional field on POST /v1/images/edits: a transparent PNG laid over the first image to guide the edit. It must match that image's dimensions after orientation and contain transparent pixels, it cannot be combined with mask, and it can be sent as a multipart file or a JSON reference.

Does markup work with GPT Image 2.5?

On paper, yes: all three descriptions of the field in OpenAI's spec name 2.5, two of them 2.5 only, and openai-python has shipped the field since v3.42.0 on October 31, 2026. OpenAI's guide, changelog, model pages and pricing page don't mention it yet, and there is no official example.

Is GPT Image 2.5 Sunburst better than Flare for editing?

For one clear change they are level: both recolored the mug and got the two-line chalk text right every run. Sunburst followed our drawn notes in 1 of 2 runs, removing the jar and adding the apple, while Flare did not, and OpenAI positions Sunburst for editing precision.

How much does GPT Image 2.5 cost?

Both bill the same rates on OpenAI's pricing page (checked November 1, 2026): $8 per 1M image input tokens, $30 for image output, $2 for cached image input, $5 for text input and $1.25 for cached text input. Batch halves the image rates, and OpenAI publishes no per-image price.

Sources checked

• Spec commit of October 28, 2026, adding markup and naming 2.6
• Spec commit of October 31, 2026, cutting 2.6 from two lines
• Spec file on main, with the leftover 2.6 line
• openai-python commit of October 29, 2026, adding markup
• OpenAI model list
• OpenAI full catalog
• GPT Image 2.5 Sunburst model page
• GPT Image 2.5 Flare model page
• Changelog
• Image generation guide
• Pricing, image section
• Hacker News comment by jjcm, September 8, 2026

Updated November 1, 2026. We will add GPT Image 2.6 when OpenAI lists its model ID, and markup when OpenAI documents it.