ChatGPT Images 2.5 shipped on 8 September with two API models: Flare for speed and Sunburst for editing precision. Six days on, every plate on this page was drawn by them from one prompt, so the review is the page you are reading.
Read the release at OpenAIReleased 8 September 2026 to every ChatGPT, ChatGPT Work, and Codex tier, and to the Images API as two models.
A direct-to-consumer product page laid out as a lookbook spread, sold in two variants, with technical-drawing plates in violet on apricot.
The product page is the shelf most freelancers bill for, and an image model sold in two variants belongs on it: one brief, two finishes, one price. This page skips the usual gallery-and-buy-box split for a lookbook spread. A hang tag carries the masthead. The product title runs full width in a high-contrast serif, the technical flat runs full width beneath it, and the two variants are sold as two cards, each holding the plate its own model drew from the same prompt, with the spec rows printed underneath. Sections open with sewn-in care labels, the ground is colour-blocked in two warm tints with no texture, and the Prompt Lab sits in a violet garment bag. Instrument Serif, a sharp display serif, sits over Schibsted Grotesk, a plain grotesque body, with Sometype Mono on every label.
OpenAI shipped ChatGPT Images 2.5 to ChatGPT, ChatGPT Work, and Codex on all tiers: sharper detail, more natural lighting and texture, subjects from reference photos that stay recognisable across new settings and styles, and generation up to 50 percent faster than Images 2.0. This page is late to it on purpose. A week of use tells a designer more than a launch-day read. Introducing ChatGPT Images 2.5.
gpt-image-2.5 Flare is the new default, higher quality than gpt-image-2 at half the latency, aimed at creator content, product experiences, and high-volume generation. gpt-image-2.5 Sunburst takes longer on purpose and is sold on tighter control across edits, for campaign creative and polished product imagery. Both are priced per token at the same rate, with the same six quality settings, so the choice is about the loop you are in, not the budget. The Flare model page.
In ChatGPT itself, Sketch lets you draw a rough layout or contour in the chat and use it as the reference for the final image. Templates start a poster, a flyer, or a product photo from a format instead of a blank prompt. Comments pin an edit request to one region of the image, and a shared image can carry its prompt so someone else runs it with their own photos. For a designer, the comment on a region is the one that changes the loop, because it is the redline you already give a retoucher.
The claim that matters for production art is that 2.5 edits only what you asked for and leaves the rest alone: swap a product, a background, or a line of copy while the subject, composition, and brand treatment hold, and keep that consistent across several turns without the image degrading. OpenAI also says the model holds a requested visual direction more firmly as the brief gets more specific, which inverts the old habit of hedging a prompt. Write the direction harder, not softer.
Every plate on this page came out of gpt-image-2.5 through the Images API, from a prompt written as art direction: technical drawing, fine violet line on flat apricot paper, blank rectangular tags, one burnt-umber mark, no text. Flare drew the jacket flat in 21 seconds and the lens section in 19. Sunburst drew the same lens from the identical prompt in 25. All three held the two-colour palette exactly, kept every tag blank, placed the single umber mark where the prompt put it, and drew dimension lines that land on real edges. The gpt-image-2 plates in our last few weeks of issues did not do that reliably; they warmed the paper toward cream and added a grain nobody asked for. Between the two variants we cannot see a gap on a line drawing that justifies the longer wait. Sunburst hatched more densely and let the focal point run to the edge of the frame; Flare kept it inside the margin with a tag on it. The precision Sunburst is sold on lives in edits across turns, and a one-shot plate does not test it.
The API pitch is anchoring: variations that stay tied to a source photo, single-element edits that preserve everything around them, and a series of on-brand assets that keeps its hierarchy as the instructions pile up. That is the exact job a build agent has when it generates a hero and three supporting images for a client page. Write the direction once, in the brief, and expect the set to read as one commission. The Sunburst model page.
Adobe has added the 2.5 models to Firefly, and Manus reports Flare running two to four times faster than gpt-image-2 with better transparent backgrounds. Our image script switched its OpenAI leg to Flare on 11 September, which is why the plates here exist; it still pins the older gpt-image-1 for alpha cutouts because we have not tested 2.5 transparency yet. When we do, the result goes in this column, not in a caption.
The variant cards with a receipt. When a product comes in two versions, do not write a comparison table. Build two identical cards, each holding a real output of its own version made from the same brief, with the same spec rows printed beneath, and a receipt under the plate that says what made it, how long it took, and what it cost. The reader compares the things themselves, not adjectives about them.
Use it for tiers, finishes, fabrics, plans, and any tool sold in a fast and a careful edition. What keeps it honest: same prompt, same size, same settings, the timing measured rather than quoted, and the weaker result shown at the same size as the stronger one. If one version has nothing to show, the page has no business selling it.
Paste it into Beaver Builder AI, v0, Claude Design, Bolt, or Lovable.
Build a direct-to-consumer product page for one product sold in two variants; call the style The Hang Tag. Semantic HTML and CSS only, no JavaScript beyond a font loader. The page is a lookbook spread, not a gallery-and-buy-box split: everything runs full width in one column on a warm apricot ground (#F5D3AE), colour-blocked by section with a paler peach (#FAE4CC), no texture, no gradient, no grain. Deep violet (#2A1541) is the ink for all text and rules; a brighter violet (#5A2DB4) is the accent for links, the one button, and a hover shadow; burnt umber (#7E4A25) is the tone for small labels, captions, and italic emphasis. Type: Instrument Serif (a sharp, high-contrast display serif) at 5.4rem on desktop and 2.4rem on mobile for the product title, and for section titles and variant names; Schibsted Grotesk (a plain grotesque) for all body copy at 19px with line-height 1.7 and a measure of about 66 characters; Sometype Mono, uppercase with 0.06em tracking, for every label, caption, and spec key. Never italicize paragraphs. Masthead: a hang tag, not a table. A 232px peach card with a 1.5px violet border, a punched hole drawn as a small ring near its top edge, and a thin string rising from the hole to the nav, rotated minus 4 degrees, carrying the brand in mono, a large serif issue number, the date, and the reading time. It rotates to level on hover over 220ms. Above the fold, in order: a mono top nav (brand left, three links right); the hang tag; the product title full width in the display serif, max 18 characters per line; a one-paragraph deck at 1.28rem; a full-width framed plate (a 3:2 technical drawing with a 1.5px violet border and a mono caption); then the variant row: a serif heading, and two equal cards side by side (stacked on mobile), each with a serif variant name, a mono variant number, a square plate made by that variant from the same brief, and a definition list of spec rows (built for, speed, price, quality, this plate) separated by hairlines at 28% violet, the last row naming the measured time in bold. Under the cards, one real button in the bright violet linking to the source, with a plain sentence beside it. Sections open with a care label: an inline box with a 1.5px violet border and an inner 1px dashed stitch line inset 5px, holding a mono index, the section title in the serif, and a mono item count. Items are separated by hairlines and carry a mono date line, a bold grotesque title, and one or two sentences. The design move section is a 7/5 grid: prose left, and on the right a framed receipt panel listing every plate on the page with its model, size, quality, seconds, and the first line of its prompt in umber italic. The prompt section is a deep violet garment bag with peach text and a bordered pre block. End with a prev/index/next rail between two violet rules and a short umber colophon. Guardrails: body never under 18px, line-height never under 1.6, no justified text, small labels at WCAG AA against the surface they sit on, no fake screenshots with readable text, no pill radius, card radius 4 to 6px, hover states on links, the button, the tag, and the cards only (150 to 220ms), and a prefers-reduced-motion guard. Images: technical drawings in fine violet line on flat apricot paper with dimension lines, blank label tags, and one umber mark, no legible text in them.
A release post is a claim. A page whose pictures came out of the release is evidence, and the receipt under each one is what makes it evidence rather than decoration. Print the receipt.