How to use GPT Image 2.5 in ChatGPT and the API

Two endpoints cover everything. Generations makes a new image from text. Edits changes an image you supply, guided by your instruction and by any reference images you attach. Both endpoints take a model string, both accept the same quality settings, and both return an image rather than text.

The model parameter is where you choose speed or precision. Send `gpt-image-2.5-flare` unless you have a reason not to. Send `gpt-image-2.5-sunburst` when an edit has to be exact and you can afford the extra wait. Pin a dated snapshot in place of the alias when output has to stay stable across months, because the alias moves under you when OpenAI updates the model.

Quality has the largest effect on both cost and wait. It runs from low through max, with auto letting the model decide for you. Higher settings spend more image output tokens, and image output is the most expensive token type in the table below. Work at a low setting while you are still changing the wording. Raise it once the composition is right and you only want a better render of a picture you already like.

Your account tier sets how much you can push through per minute. Two ceilings apply at once, a token budget and an image count, and you stop at whichever one runs out first. Low tiers hit the image ceiling more often than the token ceiling, because a handful of high quality images can exhaust the per-minute image allowance long before the tokens run down.

Establish the scene before you get to the details. The model reads outward from the context you set, so the first sentence should say what kind of picture this is and where it happens. A prompt that opens with a stray detail spends its best attention on that detail. A prompt that opens with the setting gives every later instruction somewhere to attach.

Put any text that must render exactly inside quotation marks, and say plainly what it reads. Guessing is what produces mangled lettering. Keep rendered text short as well, because short strings are where these models are most reliable. A word or two on a sign lands consistently. A full sentence on a poster is a gamble no matter how clearly you ask.

Name materials instead of asking for realism. The word realistic tells the model almost nothing, because every picture it makes is realistic by its own reckoning. Visible pores, a loose fabric weave, chipped paint, fingerprints on glass: these are instructions it can act on. Specific nouns beat general adjectives every time in this kind of writing.

Name a camera or a medium when you want a particular look. A lens length, a film stock, a printing process or a painting technique carries a whole set of conventions with it, and the model already knows those conventions. That single phrase does more work than a paragraph describing the same effect in your own words.

Say what to avoid as well as what to include. Exclusions are cheap to write and they prevent the most common surprises. No text, no watermark, no extra hands, no crowd in the background. Each one closes off a failure you would otherwise have to fix in a second pass.

Change one thing at a time when you iterate. A rewritten prompt returns a different picture, and you learn nothing about which edit mattered. This is craft advice rather than documented model behaviour, but it holds across every image model worth using, and it holds here.

Text came out wrong. The fix is almost always shorter text and firmer wording. Put the exact string in quotation marks and state that it is the text on the sign or the label. Cut it to a few words if you can, because reliability drops as the string gets longer. If the layout matters, sketch it in ChatGPT rather than describing where each word sits.

The reference subject drifted. This shows up as a face that is nearly right and a person who is nearly someone else. Move the request to the precision model, since preserving reference subjects is what it is built for. Then name the features that must survive rather than asking for the person to be kept. Hair, jaw, skin texture and clothing folds are the details that go first.

The edit changed too much. A general instruction invites a general redraw, so scope it. In ChatGPT, pin a comment on the region instead of typing into the chat. In the API, state what must stay untouched in the same prompt that states what changes. If the result is still too broad, roll back through version history and try a narrower instruction rather than editing the damaged image again.

Results drift between runs. Two causes are common. You may be calling the floating model alias while OpenAI updates the model underneath, which a dated snapshot fixes. Or your prompt may be leaving too much unsaid, in which case the model fills the gaps differently each time. Add the missing constraints and the spread narrows.