The cheapest image generation API is not the same model at every resolution. As of September 2026, GPT-Image-2 costs $0.0633 per image at 1K, 2K, and 4K, while Google's models on the same key have a separate rate for each resolution tier. So GPT-Image-2 is the most expensive generation model at 1K and the cheapest at 4K—below both Nano Banana 2 at $0.1057 and Nano Banana Pro at $0.1200. Rankings reverse within the range where comparison articles use just one number. Choose the resolution you will actually output before choosing a model.
First, our position: Kunavo is our own product, a gateway that resells access to models, not an independent benchmark. The figures below are read at page render time from the catalog that bills requests. Kunavo rates last verified: October 8, 2026.
Price per image, by model and resolution
One price per image—the unit is the output, not the prompt length. “Provider list price” is the price the provider itself publishes for the same image, verified on Google's Gemini API image generation documentation and OpenAI's API pricing page. A dash means that resolution is not available as a separate tier for that model.
| Model | Provider | 1K | 2K | 4K | Provider list price (base) |
|---|---|---|---|---|---|
| Nano Banana 2 Lite | $0.0255 | — | — | $0.0340 | |
| Nano Banana | $0.0273 | — | — | $0.0390 | |
| Nano Banana 2 | $0.0469 | $0.0707 | $0.1057 | $0.0670 | |
| GPT-Image-2 | OpenAI | $0.0633 | $0.0633 | $0.0633 | $0.1266 |
| Nano Banana Pro | $0.0670 | — | $0.1200 | $0.1340 |
Models without resolution tiers are billed at a single rate regardless of the size requested. All rates in the table are below the provider's published price for the same image.
Models billed by rendering speed rather than resolution
Ideogram v3 has no resolution tiers. Its billing dimension is rendering speed—the amount of compute used per image—with a tier structure like resolution. The cheaper setting is for drafts; the more expensive one is for a final deliverable. Keep this in mind before comparing the figures in the table above with the headline rate: the two do not measure the same thing.
| Model | Rendering speed | Per image | Provider list price |
|---|---|---|---|
| Ideogram v3 | Turbo | $0.02625 | $0.0300 |
| Ideogram v3 | Balanced | $0.0525 | $0.0600 |
| Ideogram v3 | Quality | $0.0750 | $0.0900 |
Every tier is priced below the price Ideogram publishes for the same setting. Ideogram v3 Edit is billed at the same 3 speed tiers and is listed in the editing section below—it is the only model on this key that accepts masks.
Where the ranking reverses
At 1K, comparison articles are right—smaller Google models are cheapest and GPT-Image-2 is more expensive. At 4K, the order reverses. OpenAI does not charge extra for larger outputs, while Google does. Specifically, 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 times as much as the cheapest option.
- At 4K—GPT-Image-2 $0.0633, Nano Banana 2 $0.1057, Nano Banana Pro $0.1200. GPT-Image-2 is the cheapest of the three here.
For a product generating thumbnails, Google models win on cost by a wide margin. For print-resolution assets, a single-rate model wins. Articles that name just one winner do not say which of these use cases they assume. To estimate the cost at your actual monthly volume, use the AI image generation cost calculator—switching tiers shows how the model rankings change.
Editing existing images
There are 3 editing models that accept images as input. They use the same key and balance.
| Model | Per image | Multiple input images | Mask-based inpainting |
|---|---|---|---|
| Nano Banana Edit | $0.0273 | Support | Not supported |
| GPT-Image-2 Edit | $0.0633 | Support | Not supported |
| Ideogram v3 Edit | $0.02625 onward (by speed) | Not supported | Support |
Models that accept multiple source images in 1 request: Nano Banana Edit and GPT-Image-2 Edit. This capability is needed for compositing, and edits are directed through prompts rather than painted regions. Only Ideogram v3 Edit accepts mask-based inpainting (since September 14, 2026). Choose it for workflows such as retouching, where you paint the area to change; choose the other 2 models when compositing multiple images. Each edit is also billed as 1 image, so a pipeline that generates an image and then edits it 2 times costs the equivalent of 3 images per final result.
What Kunavo does not offer
The 5 model families often shortlisted for this search are not available on Kunavo. As of September 2026, these are Flux, Seedream, Recraft, Stable Diffusion, and Midjourney. If your shortlist is built around one of these, another provider is the right choice; this page is not trying to persuade you to switch away from them. Ideogram was on this list until September 14, 2026, but it is now available with this key—see the rendering speed section above.
How to call it
Generation and editing each use a single endpoint, both in an OpenAI-compatible format. Only the model ID changes between models. Unlike video generation, image generation returns the result immediately without task polling. The request format and sizes are covered in the image endpoint documentation, with model-specific details in Nano Banana Pro pricing and how to use GPT-Image-2. For video generation on the same key, see the video generation API comparison.
Frequently asked questions
Which image generation AI API is the cheapest?
The answer depends on the output resolution. As of September 2026, on Kunavo, Nano Banana 2 Lite is the cheapest at 1K, at $0.0255 per image, while GPT-Image-2 is more expensive at $0.0633. But at 4K, GPT-Image-2 stays at $0.0633, the same price as at 1K, while Nano Banana 2 rises to $0.1057 and Nano Banana Pro rises to $0.1200, making GPT-Image-2 the cheapest of the three. That's because OpenAI does not charge by resolution, while Google does. Choose the resolution first, not the model.
How much does the image generation API cost?
As of September 2026, on Kunavo, the price ranges from $0.0255 to $0.1200 per image. The unit is one output image, not a number of tokens, so a token calculator cannot estimate the cost of this workload. Usage is billed against a prepaid balance; failed generations are not charged, and there is no monthly minimum.
Does Kunavo have Flux, Seedream, Recraft, Stable Diffusion, and Midjourney?
No. As of September 2026, Kunavo's image catalog includes the Google Nano Banana family, OpenAI GPT-Image-2, and Ideogram v3—6 generation models and 3 editing models. Flux, Seedream, Recraft, Stable Diffusion, and Midjourney is not available on Kunavo. If your work requires one of those, another provider is the right choice. Ideogram was added on September 14, 2026, and is billed by rendering speed rather than resolution, so compare prices at the speed you will actually use.
Which models can edit existing images?
As of September 2026, there are 3. Nano Banana Edit costs $0.0273 per image, GPT-Image-2 Edit costs $0.0633 per image, and Ideogram v3 Edit ranges from $0.02625 per image depending on rendering speed. They edit in different ways. Google's and OpenAI's editing models accept multiple source images in one request, and edits are specified with a prompt rather than a mask—useful for compositing product images into a scene. Only Ideogram v3 Edit accepts a mask (which specifies the painted area to change) with this key—useful for retouching workflows.
Why does the image API charge per image instead of by token?
The cost is determined by the fixed GPU work per output, not the length of the input. A 5-word prompt and a 200-word prompt both generate one image and take roughly the same rendering time, so token billing would charge for the wrong thing. In practice, image and text costs use different units, so they cannot be combined in one calculator.
Is there a free image generation API?
Not one suitable for production. Some providers offer trial credits, and some consumer apps have free tiers with daily limits, but since each generation uses real GPU time, no provider offers an unlimited free image API. The realistic floor is a few cents per image, and Kunavo's lowest production rate is under 3 cents. If you see an API claiming to be unlimited and free, first check what it actually proxies and how it handles the prompts you send.