What image generation actually costs: measured prices, and the setting that multiplies them by seven
Image generation pricing is usually quoted as a single per-image number. In practice the same model can cost several times more depending on one setting. Everything below is what we were actually billed.
Per-image cost at default settings
| Model | Cost per image | Latency | Images per run | Output |
|---|---|---|---|---|
| GPT Image 2 | $0.0174–0.0229 (~$0.019) | ~24s | 1–10 | PNG |
| Nano Banana 2 | $0.068 | ~11s | 1 only | JPEG |
| Seedream 5.0 Lite | $0.035 (flat) | ~41s | 1–4 | JPEG |
Under these conditions GPT Image 2 is cheapest and Nano Banana 2 dearest, with latency running in the opposite order. Faster costs more, straightforwardly.
Raising quality multiplies cost by nearly seven
GPT Image 2 exposes a quality parameter (auto / low / medium / high). The $0.019 above is at auto. Running the same model at 16:9 and quality=high produced this:
| Settings | Cost | Latency |
|---|---|---|
| auto / 1:1 | ~$0.019 | ~24s |
| high / 16:9 | $0.1299 | 72.3s |
Can't I just compare the per-image price?
You rarely accept the first output. What matters is how many runs it takes to get a keeper.
Budget per accepted image, not per run
You rarely accept the first output. Real cost is set by how many runs it takes to get one keeper — which is where images-per-run matters.
Nano Banana 2 emits one image per run, so seeing five candidates costs five runs and $0.34. GPT Image 2 emits up to ten per run, so five candidates cost about $0.095. The headline per-image gap is 3.6×; including selection it widens.
| To see five candidates | Runs | Total cost |
|---|---|---|
| GPT Image 2 | 1 (five at once) | ~$0.095 |
| Nano Banana 2 | 5 | $0.34 |
| Seedream 5.0 Lite | 2 (4 + 1) | $0.07 |
Resolution affects whether the call succeeds, not what it costs
Seedream 5.0 Lite is a flat $0.035 per image; raising resolution does not change the unit price. It does, however, have a floor.
GPT Image 2 has no resolution parameter at all — it is unsupported rather than ignored. Because accepted parameters differ per model, a single shared UI showing identical controls will be lying somewhere.
Defaulting to high is the expensive habit. Use it on the final image only.
How to build an estimate
- Iterate at auto quality; raise quality only for the final image
- Use the full per-run image cap — fewer runs means less cost and less waiting
- Check each model's resolution floor and unsupported parameters; a rejected call is not a pricing problem
- Keep video out of the same estimate — it is an order of magnitude apart
Frequently asked
- Which image model is cheapest?
- Of the three we measured, GPT Image 2 at about $0.019 per image. But quality=high takes it to $0.1299, so compare settings, not just models.
- Should I turn quality up?
- It costs about 6.8× and takes about 3× as long. Iterating at auto and using high only for the accepted image worked best.
- What does video generation cost?
- An order of magnitude more. Wan 3.0 measured $0.2125 for two seconds at 480p. Details are in our article on video cost.
