GPT Image 2.5 Quality: auto, high, xhigh, max Explained

To pick the right quality tier in GPT Image 2.5, keep one rule: stay on auto by default, finish real work at high, and only climb to xhigh or max when you can point at a specific defect that high did not fix. Quality is a compute setting. It is not a rating button that makes every image prettier, and it is not an adjective you put in the prompt.

On Create, both Flare and Sunburst expose auto, low, medium, high, xhigh, and max, alongside 1K, 2K, and 4K resolutions. Confusing those two axes is the most common way to waste an afternoon.

What each of the six tiers buys

When OpenAI shipped gpt-image-2.5-flare and gpt-image-2.5-sunburst on 8 September 2026, it extended the quality ladder above high for the first time: xhigh and max are new, and earlier GPT Image models stopped at high. The tiers differ in how much reasoning budget the model spends on fine regions, not in a subjective sharpness score:

Tier	Best for	What you get	What it costs
low	Thumbnails, layout checks	Fastest and cheapest; soft detail, readable composition	Blurry edges and small type
medium	Everyday social assets	Solid texture and coherence	Tiny detail not guaranteed
high	The default for finals	Noticeably firmer detail, cost still sane	Slower than medium
xhigh	Deliverables where micro detail is the point	Steadier edges and fine texture	Higher cost, return depends on the frame
max	High-stakes hero art, stubborn prompts	The largest available compute budget	Slowest, and not guaranteed better
auto	Never touching the dial	The model decides	The same prompt may land on different tiers

auto is a default, not an optimum

auto is what OpenAI lists as default, and it is what both models start on here. The upside is that you never think about it. The downside is that it is not reproducible: the same prompt may be computed at medium today and high tomorrow.

That does not matter for a one-off image. It matters a lot for a series. You think you changed only the background, and the quality tier moved with it. The fix is simple: whenever you compare two versions, pin the tier. Use auto while exploring, then lock high for anything that has to line up with a previous batch.

Quality and resolution are two axes

People treat “4K” as if it were a quality tier. It is not:

• Resolution (1K / 2K / 4K) sets output pixels. It decides how far you can crop and how large you can print.
• Quality (low to max) sets how much compute goes into the frame. It decides whether fine detail holds up.

All four combinations are useful in practice:

1. 1K + low — cheap layout sweeps.
2. 1K + high — judge detail without paying 4K wait times.
3. 2K + high — the sweet spot for most deliveries.
4. 4K + medium — you need canvas, and post-processing will handle detail.

OpenAI still flags anything above 2560×1440 as experimental, so check those files before shipping. Frame and pixel choices are covered in the size and aspect guide; this piece stays on quality.

Why the on-site defaults differ

Create does not start the two models in the same place, and that is deliberate:

• Flare starts at 1K + auto. It is the fast model with lower latency than the previous generation, so you should not spend your budget on pixels while you are still hunting for a direction.
• Sunburst starts at 2K + high. It is the precision model, so it begins at what most people would call a final setting.

The consequence: switching from Flare to Sunburst silently changes both the tier and the resolution. A comparison that only “changed the model” is therefore contaminated. To isolate the model, set quality and resolution to the same values on both sides first. The division of labor is covered in Flare vs Sunburst.

Three signals that justify xhigh

Upgrade on a trigger, not on a feeling. Only these three situations call for xhigh:

1. Small type or chart labels must be readable, and strokes start merging at high.
2. Micro detail is the selling point — fabric weave, brushed metal, skin pores, product seams that turn to mush when zoomed.
3. Edges are dirty — halos or color bleed where the product meets the background, still visible at web size.

If none of those apply, upgrading just produces the same image more slowly. And if the problem is somewhere else entirely — hand anatomy, object count, wrong words — max will not save it. That is a prompt or reference problem.

For text-heavy frames, pair this with on-image text guidance: lock the copy and placement first, then let the tier handle detail. It works better in that order.

Always downgrade and retest

OpenAI’s line that higher quality does not guarantee better output deserves to be a step in your workflow, not a disclaimer. After any upgrade:

1. Generate one frame at xhigh or max and confirm the problem area actually improved.
2. Run the same prompt back at high.
3. If both look the same at your delivery size, stay at high.

This saves more time than any prompt trick. Complex prompts can take close to two minutes at the top tiers, and one max run is long enough for three or four high runs.

Common mistakes

• Writing “4K”, “high quality”, or “masterpiece” in the prompt: parameters live in the controls, content lives in the sentence.
• Judging material at low: low is soft by design. Using it to decide whether brushed metal reads correctly guarantees the wrong answer.
• Changing two things at once: swap the tier and the prompt together and you learn nothing.
• Defaulting to max for peace of mind: double the cost and time, uncertain gain, and fewer directions tried.
• A/B testing across tiers: a comparison is only valid when everything else is held constant.

A repeatable tier workflow

Three steps, so you stop improvising:

1. Explore — Flare at 1K with low or medium. Change one prompt layer per round until subject and composition hold.
2. Finish — Sunburst at 2K with high. Fix material, edges, and text. Move to 4K only when you need the canvas.
3. Spot-fix — upgrade the same prompt to xhigh only while a named defect survives; reserve max for frames where a rework costs more than the wait.

Then run the downgrade test. What you end up with is a habit: which subjects earn high, which are fine at medium, and when max is only emotional support.

Next step

Take the image you keep re-rolling and run it once at high, then once at xhigh. If they match at delivery size, turn that judgment into a template and go back to Create to explore more directions at a lower tier. To keep those settings stable across rounds, read multi-turn edits next.