Skip to content

Mastering AI, Automation & Generative Technologies

What It Costs to Get One Usable AI Image From an API
AI Image Generators

What It Costs to Get One Usable AI Image From an API

Dataskater Lab
5 min read

With six everyday prompts, a usable image cost 0.043 to 0.067 USD on three Gemini image models: between 1.0 and 1.5 times the list price. The rate cards say 0.034 to 0.067. We rendered 18 images, kept 15 under a strict check, and divided what we spent by what we could use.

Tested 2026-11-17 with google/gemini-3.1-flash-image, google/gemini-nano-banana-2.1 and google/gemini-3.1-flash-lite-image through OpenRouter.

Cost per render and cost per kept image for three image models. Flash-image: 0.067 per render, 0.067 per kept (strict). Nano-banana-2.1: 0.036 per render, 0.043 per kept. Flash-lite-image: 0.034 per render, 0.050 per kept.


Grey is the cost of one render, purple the cost of one kept render under strict scoring. All numbers are in the table below.

The number

Strict scoring, cost per kept image: flash-image 0.067 USD (1.0 times list, 6 of 6 kept), nano-banana-2.1 0.043 (1.3 times, 5 of 6), flash-lite-image 0.050 (1.5 times, 4 of 6).

Inputs

Six prompts written for what a reader would ask an image API for, one per use: a product photo, a portrait, an illustration, a diagram, text inside an image and a landscape. One render per prompt on each of the three models: 18 renders, no retries. None failed to return an image.

ID Use Prompt, verbatim
P1 Product photo A studio product photo of a matte ceramic coffee mug with a wooden handle, on a plain light grey backdrop, soft window light from the left, shallow depth of field, no props, no text.
P2 Portrait A head-and-shoulders portrait of a woman in her sixties with short grey hair and round glasses, smiling slightly, natural daylight, neutral background, 85mm lens look.
P3 Illustration A flat vector illustration of a fox reading a book under a tree in autumn, warm orange and teal colours, simple shapes, clean outlines, no text.
P4 Infographic or diagram A simple diagram of the water cycle for a school handout: sun, sea, rising vapour, cloud, rain, river back to the sea, each step marked with a small arrow, flat colours on white.
P5 Text inside an image A poster with the exact words "OPEN LATE" in large bold letters on a red neon sign on a dark brick wall, nothing else written anywhere in the image.
P6 Landscape A wide landscape photo of a mountain lake at sunrise, low mist on the water, pine forest on the shore, snow on the far peaks, calm and realistic.
Cost per render and cost per kept image for three image models. Flash-image: 0.067 per render, 0.067 per kept (strict). Nano-banana-2.1: 0.036 per render, 0.043 per kept. Flash-lite-image: 0.034 per render, 0.050 per kept.](../article-assets/ai-image-cost-per-kept-image/chart-v2.png)
*Grey is the cost of one render, purple the cost of one kept render under strict scoring. All numbers are in the table below.*

## The number

Strict scoring, cost per kept image: flash-image 0.067 USD (1.0 times list, 6 of 6 kept), nano-banana-2.1 0.043 (1.3 times, 5 of 6), flash-lite-image 0.050 (1.5 times, 4 of 6).

## Inputs

Six prompts written for what a reader would ask an image API for, one per use: a product photo, a portrait, an illustration, a diagram, text inside an image and a landscape. One render per prompt on each of the three models: 18 renders, no retries. None failed to return an image.

| ID | Use | Prompt, verbatim |
|---|---|---|
| P1 | Product photo | A studio product photo of a matte ceramic coffee mug with a wooden handle, on a plain light grey backdrop, soft window light from the left, shallow depth of field, no props, no text. |
| P2 | Portrait | A head-and-shoulders portrait of a woman in her sixties with short grey hair and round glasses, smiling slightly, natural daylight, neutral background, 85mm lens look. |
| P3 | Illustration | A flat vector illustration of a fox reading a book under a tree in autumn, warm orange and teal colours, simple shapes, clean outlines, no text. |
| P4 | Infographic or diagram | A simple diagram of the water cycle for a school handout: sun, sea, rising vapour, cloud, rain, river back to the sea, each step marked with a small arrow, flat colours on white. |
| P5 | Text inside an image | A poster with the exact words "OPEN LATE" in large bold letters on a red neon sign on a dark brick wall, nothing else written anywhere in the image. |
| P6 | Landscape | A wide landscape photo of a mountain lake at sunrise, low mist on the water, pine forest on the shore, snow on the far peaks, calm and realistic. |

![All 18 renders in a grid: six prompts, P1 to P6, in rows, and three models in columns in the cost-table order: gemini-3.1-flash-image, gemini-nano-banana-2.1, gemini-3.1-flash-lite-image. Every tile is captioned kept or rejected with the reason. Rejected: the P1 mug from nano-banana-2.1 and from flash-lite-image, and the P5 sign from flash-lite-image (strict only).
Every render, no selection. Rows are the prompts above, columns the models in the cost table. Red caption: rejected under the strict check, with the reason and the lenient result where it differs. Green: kept.

Assumptions

  • The use is an image you would put on a page or in a deck without editing. A render is kept if it passes every check in the table below (strict). Lenient allows one small miss. A failed render is paid for and counts in the numerator.
  • We wrote the checks before the first render.
  • Cost per render is the usage tokens the API reported times the OpenRouter list price per token. We did not read the billed charge, so reconcile against your own dashboard.
Check Pass rule
Prompt adherence every named subject, count and setting is present; nothing important added
Artifacts no warped hands, faces or objects; no smears, banding or cut-off edges
Legible text when asked for P5 only: "OPEN LATE" spelled exactly and readable; elsewhere invented text or logos count as artifacts
Fit for the stated use usable as a product listing, portrait, illustration, handout diagram, poster or wide header

We wrote a file-size row into the checks first but did not apply it: the API returns PNG or JPG, and compressing is a separate step.

Cost table: AI image API price per image and per kept image

Rate card first, then what we measured. The per-image rate card prices are for a 1K render and were checked on 2026-11-17 on Google's pricing page.

Model List price per 1K image Measured per render Spend, 6 renders Kept (strict) Per kept (strict) Kept (lenient) Per kept (lenient)
gemini-3.1-flash-image $0.067 $0.067 $0.403 6 of 6 $0.067 6 of 6 $0.067
gemini-nano-banana-2.1 $0.034 $0.036 $0.215 5 of 6 $0.043 5 of 6 $0.043
gemini-3.1-flash-lite-image $0.034 $0.034 $0.202 4 of 6 $0.050 5 of 6 $0.040

Total spend $0.820 for 18 renders; 15 kept strict, 16 lenient. Measured cost sits at or just above the list price on every model. What varies is the share you keep.

The three strict rejects, by use:

  • Product photo (P1): two of three models gave a mug handle attached only at the top and ending in the air. Flash-image held up.
  • Text inside an image (P5): all three spelled "OPEN LATE" correctly. Flash-lite-image doubled several strokes: a lenient keep.
  • Portrait, illustration, diagram, landscape: all 12 renders passed. Every diagram added a title the prompt did not ask for; we did not count that as a miss.

What changes the number

The use. The failures were on prompts where one wrong detail kills the image: a handle, a letter. Loose-result prompts (landscape, illustration, portrait) went 9 for 9. If your images are mostly those, the list price is close to your price.

How strict the check is. Lenient scoring moves only flash-lite-image, from 4 to 5 kept. Pick the check before you render.

Keep rate narrows the gap. Flash-image costs nearly twice as much per render as the other two, but kept 6 of 6 against 5 and 4. At 6 renders per model that is one or two images, so it is not a ranking.

Batch pricing. Google lists a Batch tier at half the standard price, $0.0168 per 1K image for nano-banana-2.1 and flash-lite-image (checked 2026-11-17). It halves the rate-card side, not the reject rate. We did not run it.

Not tested here (checked 2026-11-17): Black Forest Labs lists FLUX 3 Image Klein at 0.009 per 1K image; OpenAI prices gpt-image-2 per token with no per-image figure.

Method and limits

  • Six prompts, one render each, per model. A keep rate of 5 of 6 has a wide margin: another six prompts could give 3 or 6. Read the table as "the list price was close to the real price on everyday prompts", not as a ranking.
  • The third model is gemini-3.1-flash-lite-image because FLUX 3 Image Klein needs OpenRouter's separate images endpoint, which this run did not use.
  • One scorer, one pass. Close calls: the two mug handles, the doubled strokes on the sign, and three dots beside the flash-image handle, which we treated as decoration. Per-render scores and reasons are in scores-v2.csv.
  • Renders are 3:2 at the API's native size; nothing was resized.

Enjoy this article?

Subscribe