GPT Image 2.5 AI Image Generator

Every image below was generated on this page on 9 September 2026 on the Flare tier at 2K, with the exact prompt shown — first take, no retouching, nothing picked from a batch. Each group tests one claim.

Text that survives at every size — menus, posters and axis labels

GPT Image 2 was already the most reliable model we test for typography, and 2.5 holds the line. We gave it a printed menu with three sections and seven prices at small type, a market poster with a headline, a schedule line and a small footer, and a bar chart with four labelled bars and a y-axis. All of it came back correctly spelled and correctly set. The chart is the surprising one: values printed above the right bars, bars proportional to their values, axis 0–30 in steps of 10, footer intact. Qwen Image 3 printed the right numbers on broken geometry in the same test; this one got the geometry too.

Flare · 2K · 2:3 · 41s

Flare · 2K · 3:4 · 41s

Flare · 2K · 4:3 · 38s

The chart is the test worth repeating with your own numbers: it is one run, and one pass is not a guarantee. What it shows is that the model reads a bar chart as data with geometry, not as a picture of a chart — the failure mode we documented on the Qwen Image 3 page did not appear here.

Edits that keep the thing you care about

Reference editing is where the provider claims the biggest gain over GPT Image 2, and it is where our results were cleanest. We generated a product shot of a tea tin, then handed that image back with one instruction: move it onto a wooden stall at a snowy night market, keep the label unchanged. The tin, its typography and its proportions came through; the lighting was rebuilt for lanterns and snow. The harder test attached two references — the tin and a portrait of a fisherman generated earlier — and asked for the man holding the tin with the label facing camera. Face, hands and label all held. Multi-reference edits run on the Flare tier: the Sunburst editing endpoint returned an upstream error in every attempt on launch day, so we route reference edits to Flare and say so.

Flare · 2K · 1:1 · 38s — the reference for both edits

Flare · 2K · 1:1 · 55s · 1 reference

Flare · 2K · 3:4 · 55s · 2 references

Reference edits took 55 seconds against 38–41 for text-to-image at the same tier and resolution. Both edits cost the same as a plain generation: the upstream does not charge per reference image, and neither do we.

Photographic detail up to 4K, and a 21:9 frame that stays composed

The "sharper" claim is easiest to check on things with fine repeating structure. Our 4K interior is 3504 by 2336 pixels — the 4K tier renders a short side of 2880 rather than a full 4096 — and the book spines, parquet and dust in the light beams hold up at 100%. The fisherman portrait shows pores, stubble and rope fibre without the waxy smoothing that reference-free portrait models fall into. The 21:9 harbour frame is the format test: the model kept the horizon low, the boats in the lower third and the mist behind the headland, instead of stretching a 16:9 composition sideways.

Flare · 4K · 3:2 · 3504×2336 · 50s

Flare · 2K · 21:9 · 2688×1152 · 39s

Resolution is the only thing that changes the price on this model: 1K, 2K and 4K are three tiers, and the tier you pick (Flare or Sunburst) costs the same at each. Reference images add nothing.

• Anyone whose image has to carry text

  Menus, posters, charts, packaging. Our seven-price menu, three-size poster and labelled bar chart all passed first take — and the chart passed on geometry, not just spelling.

• Product and e-commerce teams

  Generate the hero shot, then move the product into new scenes by attaching it as a reference. Our tin kept its label through a studio-to-night-market swap and through a two-reference composite.

• Teams upgrading from GPT Image 2

  Same prompt habits, sharper output, and a simpler bill: three resolutions instead of nine quality-by-resolution cells, no per-reference surcharge, and 2K for less than GPT Image 2's high tier at 1K.

• Campaign and brand work

  The Sunburst tier is the provider's recommendation for production assets. It costs the same as Flare here, so the only reason not to try it is speed — and our same-prompt pair is on this page for you to judge.

• Not for: batches of ten

  One image per run, no batch parameter. If you want ten candidates from one prompt, GPT Image 2 still does that; here you re-run.

Credits, reference limits and ratio counts are read from our own model registry on 9 September 2026. Speeds are our measurements on the day, single runs per cell unless stated; they will move with load.