OpenAI API Names GPT Image 2.6 in Spec; Markup Replaces Mask, Model Not Yet Callable

On October 28, 2026 at 20:56 UTC, a bot account called openai-openapi-publisher merged commit aa88004 into OpenAI's public API specification with a 27-line change: a new markup parameter for the /v1/images/edits endpoint. Near the bottom of that change, a single field description noted that markup "supports gpt-image-2.5 and gpt-image-2.6." The model name spread quickly. The parameter that is actually usable today attracted less attention.

Markup Accepts a Drawing and Treats It as an Edit Instruction

Markup accepts a transparent PNG — supplied as a file ID, HTTPS URL, or base64 data URL — that must match the source image dimensions and contain transparent pixels. Developers render their edit intent onto this overlay layer and pass it to the model alongside the source image. The model reads the annotation as an instruction: highlight the area to change and indicate how, without writing coordinate lists or verbose text descriptions of spatial regions.

Markup and the existing mask parameter cannot appear in the same request. They route to different execution paths. mask supplies a binary region — replace whatever falls inside this boundary — while markup supplies a semantic visual annotation. The distinction matters for application design: mask is precise but input-heavy; markup accepts a hand-drawn overlay and leaves interpretation to the model.

The Commit Formalizes What ChatGPT's Sketch Tool Already Does

This capability has been available in OpenAI's consumer product since September 8, 2026, when the ChatGPT application introduced a feature called Sketch alongside the Images 2.5 update. Sketch allows users to draw annotations directly inside the ChatGPT interface, and the model interprets those drawings as references for image edits. Commit aa88004 translates that consumer workflow into an API parameter, making the same annotation-as-instruction mechanism available to developers building their own image editing applications.

GPT Image 2.6 Is a Name in One Sentence, Not a Callable Model

The string "gpt-image-2.6" appears once in the OpenAI API specification — inside the markup field description — and nowhere else in OpenAI's published documentation as of October 29, 2026. It does not appear in any model enumeration. There is no API identifier, no price, and no entry for GPT Image 2.6 in OpenAI's changelog.

Secondary coverage describing the markup parameter as a "Markup Vector Editor" with JSON stroke, path, and layer fields does not correspond to anything in OpenAI's published specification. Markup is a PNG overlay input. There is no SVG layer, no stroke command, and no JSON path object in the documented schema. Those descriptions were not sourced from the specification.

Arena model testing under the codename "shooting-star-green-alpha" has circulated in coverage and been tentatively associated with GPT Image 2.6, with early user comparisons describing it as marginally less grainy than GPT Image 2.5 Sunburst. Whether that codename corresponds to 2.6 is unconfirmed by any primary source. OpenAI's pre-release specification history includes prior instances — identifiers beginning with aurora- appeared briefly in the model listing in October 2026 before disappearing — establishing that a name in the spec does not indicate an imminent release.

GPT Image 2.5 Sunburst Still Leads Editing Benchmarks as 2.6 Waits

The Artificial Analysis image editing leaderboard as of October 29, 2026 ranks GPT Image 2.5 Sunburst first with an Elo score of 1,199, followed by Microsoft's MAI-Image-2.7-Preview at 1,193 and GPT Image 2.5 Flare at 1,182. GPT Image 2.6 has no entry.

Google's recently launched Nano Banana 2.1 offers mask-based image editing through its API. The mask-versus-markup distinction separates the two approaches at the API level: Google's current editing workflow replaces specified regions; OpenAI's markup parameter reads annotated overlays as semantic direction. Whether developers favor one paradigm over the other will become clearer once markup goes live and can be tested against actual editing tasks.

Read more: Google cuts image API prices and launches Nano Banana 2.1 with mask editing

GPT Image 2.6 will become callable when its identifier appears in OpenAI's model enumeration, a price is published, and a changelog entry goes live — none of which have occurred. The markup field already lists gpt-image-2.5 as supported and could go active before 2.6 ships. The more immediate question is whether annotation-as-instruction, once API-accessible, changes how developers structure editing workflows — and whether OpenAI enables it before competitors add equivalent overlay-based input support.