GPT Image 2.5 xhigh vs max: Is 2.25x Worth Paying?
The ladder between the five tiers is not evenly spaced, and the two new rungs at the top are the steepest. Here's what the extra buys, and the one case where the answer is clearly yes.
GPT Image 2.5 added two quality tiers above the old ceiling. If you came from GPT Image 2 you had three — low, medium, high — and now there are five, with xhigh and max sitting on top. Nobody tells you which one to put in production, so the safe-feeling move is to max it and stop thinking about it.
That move costs you 2.25x per image versus xhigh, and OpenAI's own prompting guide is unusually direct about why it may buy you nothing: "A higher setting doesn't guarantee a better result for every prompt."
Short answer. Start at the highest tier that could plausibly meet your requirement, confirm the output is acceptable, then step down until it stops being acceptable — that's OpenAI's recommended order and it's the reverse of what most people do. max earns its keep on work that gets printed large or inspected at 100%. For anything that ends up on a screen at normal size, xhigh is usually where the ladder stops paying.
The ladder isn't evenly spaced, and that's the whole story
If the five tiers stepped up smoothly you could pick one by feel. They don't. Each step multiplies what the previous one costs, and the multipliers are all different:
| Step | Cost multiplier | What it means |
|---|---|---|
low → medium | 2.2x | Cheap to leave draft territory |
medium → high | 4.0x | The steepest step on the ladder |
high → xhigh | 1.8x | The gentlest step, and the one most people skip |
xhigh → max | 2.25x | The decision this page is about |
Multipliers computed by us from the token rates published on the Flare and Sunburst model pages · read 12 September 2026
Two things fall out of that table that are worth more than any quality comparison.
high → xhigh is the cheapest upgrade available — 1.8x, the smallest multiplier on the ladder. If you're sitting on high because it sounds like the top of the range, you're one of the least expensive steps away from a tier that didn't exist a generation ago.
medium → high costs more than xhigh → max. The step people take without thinking is steeper than the one they agonise over.
Why this page talks in multipliers
Ratios are what transfer. Absolute per-image cost depends on the channel you run through, the size you ask for and how you're billed, and quoting one number invites you to compare things that aren't comparable. The multiplier between two tiers is the same everywhere, and it's the only figure the decision actually needs.
OpenAI's own guide tells you not to default to the top
This is the part that surprised us, because it cuts against how every model gets marketed. The official prompting guide doesn't describe the top tier as the best one. It describes it as conditional:
Usexhighormaxonly when they improve an unmet quality requirement within your latency budget. A higher setting doesn't guarantee a better result for every prompt.OpenAI — Image prompting guide
Read that twice. Only when they improve an unmet requirement — the tier is a response to a problem you've already identified, not an insurance policy. And doesn't guarantee a better result is about as close as a vendor doc gets to saying you may pay more for the same picture.
And the selection order is the reverse of what you're doing
The same guide recommends starting high enough to meet the requirement, then testing downward once it's met. Most people do the opposite — start cheap, escalate when disappointed — which costs more attempts and tells you less, because you never find out where acceptable actually stops.
- Pick the tier you believe will clear the bar.
- Confirm it clears it on your real prompt, not a test one.
- Step down one tier. Still acceptable? Step down again.
- Stop at the first tier that fails, and ship the one above it.
Three runs settles it, and you only pay that cost once per prompt family rather than once per job.
If you came from GPT Image 2, your high is not the old high
This is the trap, and it's silent. The tier labels carried over but the budgets behind them did not.
| What you set on GPT Image 2 | Equivalent budget on 2.5 | What happens if you don't change it |
|---|---|---|
low | low | Unchanged |
medium | high | Leaving medium in place moves you down |
high | max | Leaving high in place lands you at the old medium budget |
Tier lists from the GPT Image 2 and GPT Image 2.5 model pages · the equivalence column is our reading of them, not a mapping OpenAI publishes · read 12 September 2026
So a migration that changes the model ID and leaves quality: "high" alone gets noticeably cheaper and quietly drops a rung. That's fine if you were over-provisioned. It is not fine if somebody later reports the output looks worse and nobody can explain why the code didn't change.
xhigh has no GPT Image 2 equivalent at all. It's a genuinely new rung between the old medium and old high budgets, which is why no migration guide mentions it — there's nothing to map it from.
The rest of what moved between the two generations is on what changed from GPT Image 2.
When max is worth it, by the job
| The job | Tier | Why |
|---|---|---|
| Drafts, layout tests, thumbnails | low / medium | Nobody inspects these. Spending here is spending on nothing |
| Social, blog headers, in-app imagery | high | Viewed at normal size on a screen. The ladder stops paying here for most people |
| Ecommerce main images, anything with a zoom control | xhigh | Someone will zoom. And it's the cheapest step on the ladder |
| Print, large format, client delivery, anything inspected at 100% | max | The one case where the top tier is the answer rather than a hedge |
Our recommendation, not OpenAI guidance · the step-down check it points at is OpenAI's own, in the image prompting guide
The honest summary: max is for output that gets examined, and xhigh is for output that gets looked at. Most work is the second kind.
What it costs here
On this site the tiers consume credits in the same proportions as the ladder above — max takes 2.25x the credits of xhigh, and the high → xhigh step stays the cheapest one. That's deliberate: a ratio you learn on this page should still be true at checkout.
xhigh and max are on the paid plans. low through high are available on the free tier, which is enough to run the step-down check before you decide anything. What each plan includes is on our plans and credits.
What this page doesn't claim
We have not published a tier-by-tier image comparison
Nobody has, as far as we can find — every quality-tier comparison on the web right now is about GPT Image 2, which only had three tiers and neither of the two this page is about. So treat the job table as a starting point and run the step-down check yourself. The generator at the top has the tier selector open for exactly that.
One thing we can say about what the top tiers do not fix: the fine grain some people see on skies and studio backdrops is a separate issue, and whether tier affects it is untested by anyone. That's covered, with what is and isn't known, on the noise and tint tests.
Choosing between the two models is a different decision from choosing a tier, and it comes first: Flare or Sunburst.