GPT Image 2.6 Is Named in OpenAI's Spec and Nowhere Else

GPT Image 2.6 exists this week as a string inside a field description and nothing more. On 28 October 2026, a commit to OpenAI's public OpenAPI repository added a new image-edit parameter called markup and described it with one sentence that matters: "Supports GPT Image 2.5 and 2.6." Everything else you would need in order to call a model is missing. There is no gpt-image-2.6 identifier in the spec's own enumeration, no price, no changelog entry, no leaderboard row and no announcement. The OpenAI image models the same file actually lists as callable are still gpt-image-1-mini, gpt-image-1.5, gpt-image-2, GPT Image 2.5 Sunburst and GPT Image 2.5 Flare. So this is not a launch story. It is a story about what one sentence in a vendor's own machine-readable contract tells you, and what it deliberately does not.

The evidence, down to the commit

The change is small enough to read in full. Commit aa88004, authored by the OpenAI OpenAPI Publisher bot at 20:56 UTC on 28 October, is titled "Add 'markup' property to image editing endpoints" and touches 27 lines across openapi.json and openapi.yaml. It carries a reference to API documentation source revision 1326001, which is how the generated reference knows which prose to render. The same four sentences now appear on the image-edit reference page at developers.openai.com.

Here is the whole of what markup is, according to OpenAI's own spec:

• It is a transparent PNG overlaid on the first image in the edit request.

• It can be supplied as a file ID, an HTTPS URL, or a base64 data URL.

• It must match that image's dimensions after orientation, and it must contain transparent pixels.

• It supports GPT Image 2.5 and 2.6, and it cannot be combined with mask.

That is the entire public description. No diagram, no example, no explanation of how the overlay is interpreted, no statement about what 2.6 does differently from 2.5, and no clue about timing. The spec is machine-readable and exhaustive about field shapes while being almost silent about semantics, which is normal for a contract and frustrating for anyone trying to plan around it.

What is missing is more informative than what is present

Search the 5 MB openapi.json for any identifier containing gpt-image-2.6 and you get nothing. The supported-model prose on the edit endpoint enumerates the family by name and stops at the two September SKUs and their dated snapshots: gpt-image-2.5-sunburst, gpt-image-2.5-sunburst-2026-09-08, gpt-image-2.5-flare, gpt-image-2.5-flare-2026-09-08. A model that cannot be named in a model field is not a model you can call, and the repository that would carry that name has not been updated to carry it.

This is the pattern OpenAI used ahead of GPT Image 2.5 in early September: references surface in generated documentation and specification files before the product page, the price table and the changelog catch up. It is also the pattern that produced three aurora- identifiers in OpenAI's model listing on 17 October, which appeared and were pulled again within minutes and have not been explained since. A name in a spec is a leading indicator. It is not a promise, and it is not a date.

Markup is an input, not a model — and that is the real change

The version number is the headline, but the parameter is the substance. Today an image edit can take a mask: a binary region that says "change here, leave the rest alone". The prompt has to carry all the meaning, because the mask carries none — it is a hole with no semantics inside it. Markup is a different kind of object. It is a picture laid over the picture, and a picture can carry an arrow, a box drawn around a logo, a scribble in the margin, two regions distinguished by colour, or a handwritten note. If the model reads that overlay as instruction rather than as pixels to blend, the edit request stops being "replace this region" and becomes "do what the drawing says".

That last sentence is inference, not documentation. OpenAI has published exactly four sentences about markup and none of them describe interpretation. What the spec does say unambiguously is the exclusivity: markup cannot be combined with mask. So this is not an additional input bolted onto the existing edit flow. It is a second route through the endpoint, and each request takes one or the other. Existing mask-and-prompt calls are unaffected because the field is optional, but a product that wants annotation-driven editing will be choosing a different call shape rather than extending the current one.

There is a consumer-side echo worth noting. ChatGPT Images 2.5 shipped with a Sketch tool on 8 September — draw on an image, ask for a change. Markup reads like the API-side expression of the same idea: the drawing becomes the instruction, and the prompt becomes a caption for it. Whether that is what OpenAI intends is unconfirmed, but the field's constraints — same dimensions, transparent pixels, mutually exclusive with mask — fit that reading and fit little else.

The other half of the leak is an Arena codename

The signal that started this piece pairs the documentation sighting with blind Arena testing: gpt-image-2.5-sunburst on the left, an anonymous shooting-star-green-alpha on the right, with two checkpoints tested and a second output posted in the replies. A Korean-language write-up published on 28 October describes the same codename family appearing on Arena's image leaderboard — its headline names both a Green Alpha and a Red Alpha — reports that it looked only marginally different from GPT Image 2.5 with slightly less graininess, and warns that several generations were needed before one came out well.

Nothing verifies that shooting-star-green-alpha is GPT Image 2.6. The link is circumstantial: the codename surfaced in the same week as the spec sentence, it is being compared against the current top editing model, and the -alpha suffix matches the aurora-*-alpha identifiers from 17 October. Arena codenames have been misattributed before, and two differently coloured variants are not proof of two distinct model versions — they can just as easily be two sampling checkpoints of the same one. Treat the arena half as a sighting by testers, not as a benchmark result, and treat "marginally better than 2.5" as the most likely outcome until someone publishes a scored comparison.

What 2.6 would have to beat

The relevant board is image editing, because markup is an editing feature. As displayed on Artificial Analysis on 29 October, the top of that leaderboard looks like this:

• GPT Image 2.5 Sunburst (max) — OpenAI, Elo 1,199, first.

• MAI-Image-2.7-Preview — Microsoft AI, Elo 1,193, second.

• GPT Image 2.5 Flare (max) — OpenAI, Elo 1,182, third.

• MAI-Image-2.6 — Microsoft AI, Elo 1,151, fourth, level with Krea 3 (Ultra) and Nano Banana 2.1.

• GPT Image 2 (high) — OpenAI, Elo 1,136, eighth.

There is no GPT Image 2.6 entry on that board, which is the cleanest available confirmation that nothing has shipped. There is also no room at the top: the gap between first and second is six points, and the gap between Sunburst and its own Flare sibling is seventeen. Both carry 95% intervals that overlap — 1,188 to 1,210 for Sunburst, 1,171 to 1,193 for Flare. A 2.6 that only reduces graininess, as the Arena testers describe, would land inside that noise rather than above it. The interesting question is not whether 2.6 beats 2.5 on Elo. It is whether the markup path changes what an edit request can express, because that is the kind of capability a leaderboard of paired comparisons is bad at measuring.

The price it would inherit, and what you can call today

GPT Image 2.5 launched on 8 September with two API SKUs — Flare as the default, pitched at 50% lower latency than GPT Image 2, and Sunburst as the precision-editing variant — and it kept GPT Image 2's rate card exactly:

• Image input — $8.00 per 1M tokens, cached image input $2.00 per 1M, image output $30.00 per 1M.

• Text input — $5.00 per 1M tokens, cached $1.25 per 1M.

• Batch — half price: $4.00 image input, $1.00 cached, $15.00 image output, $2.50 text input.

OpenAI has published no price for GPT Image 2.6, and there is no per-image list price on its own pricing page at all — billing is by token, so cost per generated image moves with size, quality and how much reference imagery you send. For scale, Artificial Analysis's leaderboard carries an independent estimate of $210.70 per 1,000 max-quality images, about 21 cents each; use a figure like that for budgeting, not a per-image guess someone posted. If 2.6 does launch, the useful default assumption is the current rate card, since 2.5 changed nothing from 2.0; but that is one data point, not a policy, and a price change would be the more newsworthy event.

Both 2.5 SKUs are already callable through OrcaRouter today, at $8.00 input and $30.00 output per 1M tokens — the provider's list price passed through with 0% markup, on the same key as the rest of the catalogue. GPT Image 2.6 is not on the catalogue, and nothing in this leak suggests a date on which it will be. What the routing layer does buy you is the cheap part of the transition: when a new image model does appear, the routing DSL lets you send a slice of traffic to it while the rest stays on the SKU you trust, and automatic failover covers provider errors. It cannot invent a second upstream for a model only one provider serves, so a brand-new checkpoint is still a single point of failure until it is widely mirrored.

Four signals that would turn a name into a model

If you want to know when this stops being a leak, watch the specification rather than the timeline:

• A gpt-image-2.6 identifier appearing in the spec's model enumeration, or in the edit endpoint's list of supported GPT image models. That is the moment a call becomes possible.

• The changelog or a product post from OpenAI. The company has never shipped a GPT Image generation without one, including the 2.5 release that went out across ChatGPT, ChatGPT Work and Codex on the same day.

• A scored entry on Artificial Analysis, ideally with an editing Elo and a sample count. Vendor-curated arena screenshots are not that.

• The markup field going live in the API itself rather than sitting in a spec revision — a parameter you can actually pass, with a validation error when the overlay dimensions are wrong.

There is also a falsifier. If the next revision of the spec quietly drops the "and 2.6" clause, the likely reading is that a planned SKU was renamed or delayed, exactly as the aurora- identifiers disappeared in October. Documentation previews get pulled; that is what makes them previews.

For now the practical answer is unglamorous. If you are editing images at production quality today, GPT Image 2.5 Sunburst is the model with the best independent score and a published price, and it is available now; Flare is the volume choice at the same rates. Do not rewrite an edit pipeline around a parameter that has four sentences of documentation and no callable model behind it. The name is real. The model, as of 29 October 2026, is not.