GPT-5.4 Image 2 contact sheet for Nate's Image Model Arena

OpenAI

GPT-5.4 Image 2

The most believable photographs here — for triple the time and money.

OpenAI / 1:1 / openai/gpt-5.4-image-2

40 prompts rendered by GPT-5.4 Image 2

Same standard set, same framing, same model comparison surface.

Nineteen vs. forty, through the ages

Same eleven eras, two life stages, side by side: a group of fresh-faced nineteen-year-olds doing what the young did — hunts, dances, mosh pits, raves — next to middle-aged forty-year-olds doing what the settled did — harvests, workshops, offices, backyard parties. Different people, different lives, one timeline. Watch how each model handles youth vs. age across history.

Fashion, glamour, pinup

Editorial polish, neon glamour, old-Hollywood and couture, and a 1950s centerfold — faces, skin, fabric, and styling.

Product photography

Reflections, materials, and one duck modeling everything — the commercial-shot test, with a surreal closer.

Pets in the light

Small animals, soft light — the “make-it-adorable” test.

Food & cravings

A six-patty monster, two models eating it, and a humble popsicle dressed up like luxury perfume.

Worlds & abstract

A neon-cyberpunk country town and a geometric color explosion that melts — imagination over realism.

So — how did GPT-5.4 Image 2 do?

The most believable photographs here — for triple the time and money.

~187sper image
$0.23per image
40 / 40zero refusals
$9.21the whole set

The realism winner

Its people look like actual photographs. Natural skin, real un-airbrushed faces, believable mall-and-garage light — the 1990 food court could pass for a film still. If the bar is “could this be a real photo,” GPT clears it highest.

Youth vs. age

The most convincing ages. Its nineteen- and forty-year-olds both look like real photographs of those ages, in fully distinct life-stage scenes — the realism makes the youth/age contrast land hardest. (Text still weak.)

Can’t letter either

Like the base Nano and Seedream, in-image text is weak — the toolbox comes back blank, no “MISS JULY.” Grok and Nano Pro are still the only two here that actually spell.

The long pole

Slow and expensive. ~$0.23/image, $9.21 for the set, and 30+ minutes — roughly 5–6× the cheap models. You pay for the realism in both time and money.

No refusals this time

Its filter never tripped on this clothed, tasteful battery, so the full 40 rendered. On spicier prompts it remains the model most likely to decline outright.