Cheapest AI image generation API (2026) — price per image at 1K, 2K and 4K, and why resolution reorders it

The cheapest image generation API is not the same one at every resolution. As of September 2026, GPT-Image-2 bills a flat $0.0633 an image at 1K, 2K and 4K, while every Google model on the same key prices by resolution — so GPT-Image-2 is the dearest generator at 1K and the cheapest at 4K, where it undercuts Nano Banana 2 at $0.1057 and Nano Banana Pro at $0.1200. The ranking inverts across the range most comparison pages quote a single number for. Decide the resolution you actually ship before you compare models at all.

Bias declared: Kunavo is our product — a resold-access gateway, not an independent benchmark. The figures below are read live from the catalog that bills the requests, so they are what you would be charged rather than a rate copied from a launch post. Kunavo carries five image generation models and two edit models; the section on what we do not carry names the six families that are not on this key, which is most of what people shortlist.

Price per image, by model and resolution

One image, one price — the unit is the output, not the prompt length. Kunavo rates last verified September 14, 2026. "Provider list" is the vendor's own published rate for the same image: Google's, from its Gemini API image-generation documentation and pricing page, and OpenAI's, from its API pricing page. A dash means the model does not offer that resolution as a separate tier.

Models without separate resolution tiers bill their single rate whatever size you ask for. Every rate in the table sits under the provider's own published list price for the same image.

The family that prices by render speed instead

Ideogram v3 does not have resolution tiers at all. It bills by render speed — how much compute it spends reaching the image — which is a different lever with the same shape: the cheap setting is for drafts and the expensive one for the frame you ship. Worth knowing before you compare its headline rate against anything in the table above, because the two numbers are not measuring the same thing.

Every tier sits under Ideogram's own published rate for the same setting. Ideogram v3 Edit prices on the same three-speed scale and appears in the editing section below — it is the only model on this key that accepts a mask.

The resolution inversion, stated plainly

At 1K, the ranking is what every roundup prints: the small Google models are cheapest and GPT-Image-2 is the expensive one. At 4K the order reverses, because OpenAI does not charge more for the larger output and Google does. Concretely, as of September 2026 on Kunavo:

• At 1K — Nano Banana 2 Lite $0.0255, GPT-Image-2 $0.0633. GPT-Image-2 costs about 2.5× the cheapest option.
• At 4K — GPT-Image-2 $0.0633, Nano Banana 2 $0.1057, Nano Banana Pro $0.1200. GPT-Image-2 is now the cheapest of the three.

If your product renders thumbnails, the Google models win on cost by a wide margin. If it renders print-resolution assets, the flat-rate model wins. A page that gives you one winner has assumed one of those two workloads without telling you which.

Editing an existing image

Three models take an image in as well as out, on the same key and balance as generation — and they do not all edit the same way:

The two Google and OpenAI editors take more than one source image in a single request, which is what compositing needs, and direct the edit entirely by prompt. Ideogram v3 Edit is the one model here that accepts a mask — a painted region that confines the change — which is the difference that decides a retouching workflow. If you need masks, that is the row; if you need several source images composited, it is one of the other two.

What Kunavo does not carry

Five families that appear on most shortlists for this search are not available on Kunavo as of September 2026: Flux, Seedream, Recraft, Stable Diffusion and Midjourney. If your shortlist is built around one of them, a different provider is the correct answer and this page is not trying to move you off it. Ideogram was on this list until September 14, 2026 and is now on the key — see the render-speed section above.

Calling it

Generation and editing are two endpoints, both OpenAI-compatible in shape, and the model id is the only thing that changes between models. Unlike video, image generation returns inline rather than as a polled task.

# One endpoint for every image model; the model id is the only change.
curl https://api.kunavo.com/v1/images/generations \
  -H "Authorization: Bearer $KUNAVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-2-lite",
    "prompt": "a brass compass on a folded paper map, overhead, soft window light",
    "size": "1024x1024"
  }'

# Editing an existing image is a different endpoint, same key and balance.
curl https://api.kunavo.com/v1/images/edits \
  -H "Authorization: Bearer $KUNAVO_API_KEY" \
  -F model=nano-banana-edit \
  -F 'prompt=replace the background with a plain studio grey' \
  -F image=@photo.png

Per-model detail lives on the individual guides: Nano Banana, Nano Banana Pro, GPT-Image-2, and the images endpoint reference. For the video half of the same key, see the video generation API comparison.

FAQ