GPT Image 2.5: the quality knob is the expensive one, not 4K

#openai #api #ai #productivity

I assumed 4K was the expensive thing.

Then I read the price matrix properly: at 1K, max quality is 13 credits. At 4K, low quality is 8. The "bigger" image is the cheaper one, and the knob I thought was free — quality — is the one that runs the bill up by 4.3×.

I work on Callirra, where GPT Image 2.5 is selectable with both axes exposed, so the table below is our published price list rather than an estimate.

GPT Image 2.5 exposes two independent output axes: resolution (1K, 2K, 4K) and quality (low, medium, high, xhigh, max). Every combination has its own price. Most people move one axis and assume the other is a rounding error.

The full matrix

Resolution	low	medium	high	xhigh	max
1K	3	4	6	9	13
2K	5	7	10	15	22
4K	8	11	16	24	35

Credits, one credit = $0.01. The default when you do not touch anything is 1K + high = 6.

Three readings of the same table:

Quality costs 4.3× within one resolution (3 → 13 at 1K).
Resolution costs 2.7× within one quality tier (3 → 8 at low).
The cheapest 4K is cheaper than the most expensive 1K (8 vs 13).

That last line is the whole post. If you have been treating "4K" as the premium option and "low" as the free option, the pricing says otherwise.

What each axis actually buys

Resolution buys how much picture there is. Pixels, crop headroom, print size, how much a layout can be resized before it falls apart. A 4K frame is 3840-ish on the long edge depending on ratio; a 1K frame is a quarter of that in area.

Quality buys how the picture is built. The model's internal effort: more passes, better text placement, cleaner edges, more detail in the parts that matter. At max, typography is usually exact and small type survives; at low, it is a sketch that happens to be a full image.

They are not interchangeable, and the matrix prices them accordingly.

The decision rule I now use

What the deliverable needs	Buy
Crop headroom, print, a big layout that gets resized	4K + low (8) or 4K + medium (11)
Exact type, a poster, an interface, a diagram	1K or 2K + xhigh/max
A social image nobody will zoom	1K + low/medium (3–4)
A hero shot that gets judged at 100%	2K + high (10) is usually the ceiling

The trap is the middle: 2K + xhigh (15) is more expensive than 4K + medium (11) and buys you neither the pixels nor the typography.

How it looks in the request

curl https://api.callirra.com/v1/images \
  -H "Authorization: Bearer $CALLIRRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2-5-sunburst",
    "prompt": "an editorial launch poster…",
    "resolution": "2K",
    "quality": "high",
    "aspect_ratio": "2:3"
  }'

Both fields are separate, both are quoted before the run, and the quote is the number in the table above — not an estimate. The rest of the model's surface: up to 16 reference images on image-to-image, prompts up to 20,000 characters, and eleven aspect ratios including 21:9.

Where this goes wrong

Paying for max on a prompt that cannot use it. max at 1K is 13 credits; if the composition is loose and the type is one word, you paid for detail the image does not contain.

Paying for 4K to get sharper text. Resolution does not fix typography; quality does. 4K + low will give you a big image with worse letterforms than 1K + high.

Forgetting that one credit is a cent until you batch. Thirty product shots at 4K + xhigh is 720 credits — $7.20 for one product set, and more than half of a Starter month's 800.

Assuming the quality knob is free because it is a dropdown. It is the most expensive control on the page.

The same logic, cheaper

If you are iterating on composition, 1K + low at 3 credits is the cheapest full-resolution draft in the model — you can throw away four of them for the price of one 1K + max. Decide the layout at 3 credits, then spend 10–16 credits on the version you keep. That is a draft-then-finish pipeline inside a single image model, and it costs less than one "safe" mid-tier render.

FAQ

Is 4K + low really usable? For crop headroom and large layouts, yes — the pixels are real. For type and micro-detail, no.

Does quality affect generation time? Yes. In our measured runs the same prompt took roughly 24–35 s on one variant and 46–57 s on the other at xhigh. Plan batches accordingly.

What is the cheapest usable image? 1K + low, 3 credits, which is also six images in a new account's first week.

Can I change quality after a render? No. It is a generation-time setting; a re-render is a new charge.

The full matrix and the quote for every pairing are on Callirra's GPT Image 2.5 page. If you have a deliverable that genuinely needs both axes at once, I would like to see it — that is the case where the two-dimensional pricing stops being a trick and starts being a decision.