GPT Image 2.5 4K at low costs less than 1024 at medium on Sume
A 3840x2160 GPT Image 2.5 image at low quality quotes $0.0140 on Sume, below the $0.0165 for 1024x1024 at medium. Quality tier, not pixels, drives the price.

On Sume, a GPT Image 2.5 image at 3840x2160 and low quality quotes $0.0140, less than the $0.0165 for a 1024x1024 image at medium. A 4K frame has 7.9 times the pixels, yet it is the cheaper line item, because the quality tier moves the price more than the size does.
If your only reason for using a small size was cost, that reason does not hold at low and medium.
The grid
Rows are quality tiers and columns are sizes, all as Sume quotes with a one-character prompt. Cells where the larger size is cheaper than a smaller size at the next tier up are the useful ones: compare each low cell with the medium cell in the first column.
| Quality | 1024x1024 | 2560x1440 | 3840x2160 |
|---|---|---|---|
| low | $0.0074 | $0.0077 | $0.0140 |
| medium | $0.0165 | $0.0180 | $0.0325 |
| high | $0.0659 | $0.0691 | $0.1251 |
| xhigh | $0.1171 | $0.1229 | $0.2225 |
| max | $0.2635 | $0.2765 | $0.5004 |
Reading the grid
Read down a column and the price climbs by orders of magnitude: from $0.0074 to $0.2635 at 1024x1024, a factor of 36. Read across a row and it rises by a few percent to a few times: $0.0074 to $0.0140 at low, a factor of 1.9, and $0.0659 to $0.1251 at high, a factor of 1.9.
That is why a 2560x1440 low image at $0.0077 costs 5% more than the square at the same tier, and why it is far below the medium square. The estimator multiplies a quality grid by a size term, and the quality grid goes from 16 at low to 96 at max.
When to use it
Take advantage of the pattern in three situations.
- Layout tests. Render at 3840x2160 and
lowto see how the composition holds at full size for about a cent and a half. - Print and display proofs. A
lowrender at full size tells you about the crop and the margins before you pay forhigh. - Fixed budgets. If the budget is set per image at about two cents, a 4K
lowimage fits, while amediumimage at 4K ($0.0325) does not.
Limits
Do not read the grid as a recommendation to ship low at 4K. A lower tier trades detail for price, and extra pixels do not restore detail that the tier did not render. Check a sample against your standard first.
The size has to be sent as pixels, such as 3840x2160, with both edges a multiple of 16, a longest edge up to 3840, a ratio up to 3:1 and a total up to 8,294,400 pixels. The upper bound is what the quote reserves when you omit the size, which is far higher than any figure in the grid.
Large high renders are the exception to the cheap-4K pattern: $0.1251 at 3840x2160 is 1.9 times the square. Use the grid to decide where to spend. Layout work at low is nearly free at any size, medium at 4K costs about twice the square, and high at 4K is the row to approve only for final art.
A quick way to apply this is to write the tier into the stage of the workflow, not the person. Drafts and layout checks are low, review copies are medium, and anything that leaves the building is high after a person has approved the low or medium version. That rule keeps the larger sizes, which look expensive, from being the place where budgets are blown, since the tier does that. Put the same rule in your code review checklist, and when a request names high or max, ask which stage of the workflow it belongs to. If nobody can answer, it should be low.
Sources
Related posts
More in Pricing
- GPT Image 2.5 on Sume: each reference adds about 27% to the quote
Each reference image adds about 27% of the output price to a GPT Image 2.5 call on Sume: $0.0165 at medium rises to $0.0209 with one and $0.0868 with 16.
- GPT Image 2.5 with no image_size quotes 3.4x the 1024 price: set one
If you omit image_size on GPT Image 2.5, Sume quotes the upper bound: $0.2225 at high against $0.0659 for 1024x1024, 3.4x. Set the size as pixels. Tier table.
- GPT Image 2 medium costs 4x GPT Image 2.5 medium; low is the same
On Sume, GPT Image 2 costs $0.0664 at medium and $0.2639 at high, about 4x GPT Image 2.5. At low the two are within 3%. Tier table at 1024x1024.
- Estimating a gpt-realtime-2.1 call from $32 / $64 per M audio tokens
gpt-realtime-2.1 lists audio input at $32, cached input at $0.40 and audio output at $64 per million tokens. Here is the formula and a worked example.
Written by Sume