Answer: compare the same output specification, on the same date, and normalise three things first — the resolution behind the label, whether the price includes a quality axis, and whether references are charged per image. Skip any of the three and a cheaper model can cost more per delivered image.
On Callirra image pricing looks like one number per image. It is usually three numbers wearing a coat: pixels, quality effort, and inputs. Here is the checklist I use before quoting anything.
1. What resolution is behind the label?
Pixels are the first place a comparison leaks. Two models can both advertise a tier and not return the same raster:
- One model's "1080P" may return 1920×1056 — a pixel-area cap, not a rounding error.
- A "4K" tier on another may be 1920×1088 for a specific mode.
- A "2K" image may be 2560×1440 in one product and 2048-class in another.
Always compare the returned pixel dimensions, not the tier name. If a provider does not publish them, that absence is itself information.
2. Is there a quality axis hidden inside the price?
Some models have one price per resolution. Others split quality:
| Model | Price shape | Cheapest cell | Most expensive |
|---|---|---|---|
| Nano Banana 2.1 | resolution only | 3 (1K) | 5 (4K) |
| GPT Image 2.5 | resolution × quality | 3 (1K low) | 35 (4K max) |
| MiniMax H3 (video) | resolution, per second | 1 (480P) | 10 (2K) |
A model with a 12× spread between its cheapest and most expensive cell cannot be compared to a flat-price model using one number. "GPT Image 2.5 is 3 credits" and "GPT Image 2.5 is 35 credits" are both true statements about different products.
3. Are references charged?
Inputs are the third hidden line item:
- Nano Banana 2.1: up to 14 references, priced into the render, and its model page lists the caps.
- GPT Image 2.5: up to 16 references on image-to-image, same per-image price.
- MiniMax H3: the first five reference images are free; on the 2K lane each additional image is 3 credits.
- Audio: free on H3, both as input and as the built-in track.
A workflow that leans on references is more expensive on the model that meters them — and that difference does not appear in a per-image headline.
The five questions
- Pixels: what raster comes back at the setting I will actually use?
- Quality: does this price include the quality level I need, or is that a separate axis?
- Inputs: are references, audio or retries metered?
- Failure: does a refused or failed render cost anything? (On H3 and GPT Image 2.5 workflows here, failures release the charge.)
- Date: when was the number checked? Prices move monthly.
If a comparison cannot answer all five, it is a marketing number, not a price.
Cost per usable image, one more time
The arithmetic that ends most arguments:
cost per usable image = (price × attempts) ÷ images kept| Price per render | 1 in 2 kept | 1 in 3 | 1 in 4 |
|---|---|---|---|
| 3 credits | 6 | 9 | 12 |
| 6 credits | 12 | 18 | 24 |
| 10 credits | 20 | 30 | 40 |
| 35 credits | 70 | 105 | 140 |
A 3-credit render at a 1-in-4 hit rate costs 12 credits per usable image — the same as a 6-credit render that lands in one take. That single comparison is worth more than any speed benchmark, because it includes your prompt quality, not just the provider's.
What this checklist cannot fix
- Taste. A cheaper model that produces the wrong look is not cheaper.
- Latency. If a render takes four minutes, the wall-clock cost may exceed the credit cost.
- Terms. A model that refuses categories of content your brief needs is unusable at any price.
FAQ
Is per-megapixel pricing better? It is more honest than per-image, and it still hides quality effort and inputs.
Should I compare subscription prices? Only if the volumes match. List rates are the universal unit; subscriptions reward specific usage patterns.
Why do advertised rates differ from what I am charged? Because the quote depends on settings — duration, resolution, references. The number before you press generate is the only one that matters.
How often do prices change? Often enough that a comparison older than a month deserves re-checking before it goes in a client estimate.
The model-by-model pages here publish both the price and the returned dimensions, and the live rate card is at Callirra's credit-cost page. If you have a comparison where the cheaper model won per delivered image, that is the exception worth publishing — most of the time the expensive-looking model is the one that lands.