GPT Image 2.5 for Dating Photos: What It Actually Does (2026)

OpenAI released ChatGPT Images 2.5 on 8 September 2026, with two models in the API: gpt-image-2.5-flare and gpt-image-2.5-sunburst (OpenAI model reference, checked 9 September 2026). If you are looking at your dating profile and wondering whether this is finally the model that makes an AI photo that still looks like you, that is the right question, and it is not the one a launch post answers.

So the day after launch, The Ultimate Profile team ran it: four generations, two models, two modes, one real selfie as the input, then a feature-by-feature comparison of the output against the source. This is a spot check with one generation per cell, not a benchmark, and every number below is written that way. Treat it as a first-look review of what the model does to photos of people, not a verdict on every face.

From the author · freeWant to see yours?Send one selfie and we email the photo back - free, no card.Your photo is emailed when it is ready.

What GPT Image 2.5 is, and what changed since GPT Image 2

GPT Image 2.5 is OpenAI's current image generation and editing model, and it reached ChatGPT and the API on the same day. OpenAI says it is "rolling out today to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web" (OpenAI announcement, checked 9 September 2026).

In the API it arrives as two variants. OpenAI describes Flare as bringing "the same quality, editing, and speed improvements" with 50 percent lower latency than GPT Image 2, and Sunburst as adding "precision for detailed creative work, with longer generation times" (same source, same date). They bill at identical token rates, so the practical split is latency on editing rather than budget. Our companion piece on choosing between Sunburst and Flare works that decision through properly.

GPT Image 2.5 vs GPT Image 2, from OpenAI's own pages:

What	GPT Image 2	GPT Image 2.5
Quality levels	up to high	low to xhigh, max, auto
Latency claim	baseline	up to 50 percent lower (Flare)
Output image tokens	$30 per million	$30 per million
Editing	supported	stability across repeated edits

(OpenAI image generation guide plus both model reference pages, checked 9 September 2026.)

Aspect ratios matter here, because a dating app crops hard. OpenAI's guide recommends 1024x1024, 1536x1024 and 1024x1536, and accepts custom sizes with both edges a multiple of 16, a ratio between 1:3 and 3:1, and no edge over 3840 pixels (checked 9 September 2026). Portrait, which is what every dating app and every headshot crop wants, sits well inside that.

How we tested GPT Image 2.5 on a real profile photo

We ran four generations on 9 September 2026: both models, in text-to-image and image-to-image, all at 1K resolution and 3:4 portrait, launched in parallel so they shared queue conditions. One generation per cell. Four images total.

Access was through the kie.ai reseller API rather than OpenAI directly, so our wall-clock timings include that provider's queue time and are not a measurement of OpenAI's own API latency. On the editing path, Flare finished in 84 seconds against Sunburst's 185 seconds, roughly 2.2 times longer, on one run each. On text-to-image the two were 7 seconds apart, which at a single sample is noise.

The image-to-image input was our own published before-and-after asset: a real, consented, already-public 1:1 indoor selfie under flat window light, subject looking slightly downward, full beard, black t-shirt. Not a flattering photo. That was the point. The prompt constrained the model to lighting and setting only:

Keep this exact person's face, bone structure, hairline and identity completely unchanged. Restyle only the setting and lighting: place them at an outdoor cafe table in warm late-afternoon golden-hour light, natural side lighting, shallow depth of field, candid unposed dating-app profile photo, preserve authentic skin texture and pores, no beautification, no reshaping of facial features.

One structural point decides everything else here. In text-to-image, GPT Image 2.5 produced two convincing photographs of men who do not exist: photoreal at profile size, no waxy skin, no mangled ears, no broken background geometry. Neither is you. Only image-to-image, where your own photo is the reference, produces a picture of the person who turns up to the date. Our guide to AI profile picture generators covers how those model families differ; here, the run that matters is the edit.

Does GPT Image 2.5 keep your face the same?

On this one photo, yes, to a degree that surprised us. We compared the edit against the source feature by feature rather than glancing at it, which is the only way to catch likeness drift.

Feature in the source selfie	After the edit
Hairline and front cowlick	unchanged, same parting pattern
Beard density and jaw edge	unchanged, moustache transition intact
Eyebrow shape, including asymmetry	unchanged, asymmetry preserved
Nose bridge, tip and mole placement	unchanged
Face width and bone structure	unchanged

What changed was what we asked to change. Flat indoor window light became late-afternoon golden hour with a physically consistent warm rim along the hair, ear and jaw. A domestic interior became a cafe terrace with plausible depth-of-field falloff, and the 1:1 crop was extended into a 3:4 frame with invented shoulder content that holds up.

Sunburst's edit of the same source preserved the same feature set and held geometry slightly more literally: nose profile, eye spacing and brow shape sit closer to the original. It also kept more realistic skin, including visible pore detail across the forehead, forehead lines, and the natural redness on the nose tip that Flare's warmer grade partly washed out. Neither output smoothed skin into plastic. Both kept beard stubble transitions.

That is one subject, one photograph, two edits, not a claim about every face, which is why it is worth understanding how face drift happens and how to spot it before you trust any output, including this one. But if your question was whether the current generation can relight you without turning you into a better-looking stranger, on this evidence it can.

What it restyles, what it leaves alone, and what still shows

Here is the most useful thing we learned. GPT Image 2.5 restyles light and scene. It does not restyle pose and expression.

In both edits, the subject is still looking slightly downward with the same flat expression as the source selfie. Neither variant volunteered a fix. A weak input pose produces a beautifully lit photograph of a weak input pose. The light is genuinely better. The photograph is still of someone glancing down at their phone.

The consequence is blunt: the selfie you feed it matters more than the prompt you write. Get an input where your eyes are up and toward the lens, your jaw is not compressed by a downward camera angle, and your expression is one you would actually make at a table. The model will handle the golden hour. It will not handle your chin. The prompt side is the subject of our dating photo prompt guide.

Two things in the Sunburst edit are worth checking for in your own output. It added a prop nobody asked for, a glass of beer in the foreground; editing models invent scene content, and an invented prop tells a story about you that you did not choose. And a background chalkboard reads at a glance as a menu, but the lettering is not real words. At profile-picture size nobody would notice. At full size it is visible. Faces and light are excellent in this generation; incidental background text is not. Zoom to 100 percent before you upload.

As for whether a viewer can tell: OpenAI attaches signed C2PA metadata and a SynthID watermark to generated images, and is careful about the limits. "Editing, converting, or sharing a file can remove its metadata", while a SynthID watermark "is part of the image or audio itself and may survive some transformations" (OpenAI content provenance guide, checked 9 September 2026). Whether a given app permits AI-assisted photos is a separate question with different answers per platform, covered in our breakdown of dating app AI photo policies.

What GPT Image 2.5 costs, and whether it is free

Three routes reach the same model at different prices, which is why "how much does GPT Image 2.5 cost" depends on which door you use.

Route	What you pay	Checked
ChatGPT app	included in your plan, rolled out to all tiers	9 Sep 2026
OpenAI API, both variants	$5 / $8 / $30 per million text-input, image-input, image-output tokens	9 Sep 2026
kie.ai reseller, the route we tested	$0.03 per image at 1K, same for both variants	9 Sep 2026

OpenAI's figures come from the Sunburst and Flare reference pages. The API price is per token, not per image, so a reseller's per-image figure is not a quote for OpenAI's pricing.

Is GPT Image 2.5 free? Inside ChatGPT it is included in the plan you already have, and OpenAI says it rolled out across all tiers, without publishing a per-plan image quota. Expect limits and expect them to move. Through the API it is not free at any tier.

One figure matters for portraits: on the route we used, a request accepts up to 16 reference images. You do not need 16. Our edit used one photograph, which is the honest answer to how many photos GPT Image 2.5 needs for a restyle. Extra references help when holding one look across a set, not when relighting a single frame.

Where The Ultimate Profile fits

We sell an AI photo product and we use GPT Image 2.5 Sunburst in production, so treat this section as interested and the rest of the article as evidence you can check yourself.

Our position on the model matches the test. It is strong at the job the test measured, and it is not a substitute for a usable input. The Ultimate Profile turns your own selfie into photographs of you in about five minutes, no credit card required for the free tier, and uploaded photos are deleted from our servers after 30 days. What we cannot do, and what no product built on this model can do, is give you an expression you did not make. That part stays yours.

For newer reference-based options, read our guide to Nano Banana 2.1 profile photos and our guide to Seedream 5.0 Pro profile photos. Both include practical prompts and a disclosed, single-attempt comparison using the same real stock-photo references.