The Cheapest Image Generation API, With the Arithmetic

$0.0238 buys one image on our API. That is the floor, checked against our live catalog on 26 August 2026, and a thousand images at that rate comes to $23.80.

Now the awkward part about leading with that number. Cheapest per image and cheapest per usable image are different questions, and only the second one shows up on the invoice. Regenerate a $0.0238 image four times because the headline came out as smudged letterforms and you have spent $0.0952 to get one asset. Ship a $0.075 image on the first attempt and you spent $0.075. The cheap one lost.

So this post does both halves. The raw table first, because that is what you came for. Then the arithmetic that decides which row you should actually be sitting on.

Every image model we run, and what one request costs

These are per-request prices in USD, from our live catalog on 26 August 2026. The E2X column is what a customer pays; the discount against our upstream source is already applied.

Prices and product terms can change. Check each provider's current pricing before making a purchasing decision. Google Batch or Flex pricing is not directly equivalent to a standard on-demand API request because scheduling, availability, and processing conditions differ; it is therefore excluded from this comparison. This is a scoped comparison, not a claim that E2X is the world's cheapest option in every configuration.

One thing that surprises people: editing costs exactly what generating costs. Nano Banana 2 Lite edit is $0.0238, same as text-to-image. Nano Banana 2 edit is $0.04. There is no surcharge for handing the model a reference image, so an edit-heavy pipeline is not a more expensive pipeline.

Also note the legacy row. Nano Banana costs 31% more than Nano Banana 2 Lite, is a full generation older, and Google switches it off on 2 October 2026. We covered that migration in what Nano Banana actually is. Nobody should be starting a new integration there.

The multiplier that wrecks the budget

Our raw catalog number for GPT Image 2 is $0.0075. Nobody pays $0.0075.

quality is a multiplier of ×7, and it applies at every tier — low, medium, high and auto all multiply by seven. The default is high. So a standard 1K request costs $0.0075 × 7 = $0.0525, and that is the number to plan around.

This trap is not unique to us. Anyone quoting GPT Image 2 at sub-cent pricing is quoting the base before the multiplier, and a team that budgets from it lands on a bill seven times the size of the spreadsheet. We would rather you hear it here than in the third week of a pilot.

Resolution stacks on top: 2K is ×2 of the $0.0525 ($0.105) and 4K is ×3 ($0.1575).

Resolution multipliers, and the one that works in your favour

Every price above is the 1K price. Going bigger is a multiplier, and the multipliers are not the same shape across models:

Read the Pro row twice. 1K and 2K cost the same on Nano Banana Pro. If you are already paying $0.075 and asking for 1K output, you are leaving four times the pixel count on the table for nothing. Ask for 2K. There is no argument for not asking for 2K.

Two schema details that cost people money in a different way. Nano Banana 2 Lite has no resolution parameter at all. Sending one is a hard error rather than a silently ignored field, and a job that errors is a job you submit twice. And the text-to-image slug for Nano Banana 2 is google/nano-banana-2 with no capability suffix; appending /text-to-image there gives you a 404 on a request that looks perfectly reasonable.

Cheapest per image is not cheapest per usable image

Here is the arithmetic that should actually drive the decision.

Nano Banana 2 Lite is $0.0238. Nano Banana Pro is $0.075. That is a 3.15× gap, which sounds enormous until you run it as attempts:

• One Lite attempt: $0.0238
• Two: $0.0476
• Three: $0.0714 — still under Pro
• Four: $0.0952 — now Pro was cheaper

So the break-even is somewhere between the third and fourth regeneration. If a prompt lands on Lite within three tries, Lite wins and it is not close. If it routinely takes four or more, you are paying premium prices for a budget model and getting a worse asset at the end of it.

That single fact reframes the whole question. You are not choosing a price. You are choosing a hit rate, and the price follows from it.

Where the cheap tier genuinely wins

We do not want to talk anyone up a tier they do not need, so here is where $0.0238 is simply the correct answer:

• High volume, simple subjects. Textures, backgrounds, abstract fills, mood plates. Things with no small detail to get wrong.
• Social crops at 1K. A feed image is going to be viewed at 1080px on a phone. Nothing above 1K survives the platform's own recompression anyway.
• Banners and rails. Lite is the only model we carry with 1:4, 4:1, 1:8 and 8:1 aspect ratios. A 8:1 header from any other tier means generating wide and cropping, which is a second step and a worse composition.
• Anything you generate ten of and pick one from. Ten Lite attempts is $0.238. Cheap enough that selection, rather than prompting, becomes the workflow.
• Draft and comp stages. Nobody needs 4K to decide whether a layout works.
• Programmatic variation. Per-user, per-listing, per-product images where volume is the whole point and no single asset carries much weight.

For most of what most teams generate, that list covers it. Start on Nano Banana 2 Lite and only move when something specific breaks.

Where cheap quietly costs more

Four failure modes, and they are predictable enough that you can route around them before the retries start:

Text inside the image. Lite and the legacy model cannot render legible in-image text — that capability arrived with the Gemini 3 generation. A poster headline, a packaging mockup, a chart label: these are not "try a better prompt" problems, they are capability problems. Every attempt fails the same way. This is the single most common reason a cheap pipeline turns expensive, and the fix is Nano Banana Pro, which handles styled and multilingual text as a first-class feature.

Multi-reference composites. Pro takes 14 references split by role: up to 5 character images for identity, up to 6 object images for product fidelity, up to 3 style references. One ad composite, one request. Reproducing that by chaining cheap edits costs more in requests and still drifts.

Print resolution. Lite is 1K and there is no parameter to change that. If the asset ends up on paper, you need Nano Banana 2 at 4K ($0.08) or Pro at 4K ($0.15). Upscaling a 1K generation is a different job with a different failure mode.

Transparency. If you need a cut-out with an alpha channel, GPT Image 2 and its background: transparent option is the only model on our catalog that does it in one call. None of the Google tiers expose it. $0.0525 beats $0.0238 plus a background-removal step every time.

What we would actually pick

Most people asking us for Pro need Nano Banana 2. Most people asking for Nano Banana 2 need Lite. We say that a lot, and we mean it — the premium tier exists for role-split references and in-image typography, and if your job has neither of those you are buying a capability you will never call.

The routing rule we would give a team starting today: default everything to Lite, promote to Nano Banana 2 when you need above 1K or more than a handful of references, promote to Pro only when there are words in the picture or a composite that has to hold a person's face. Then measure your retry rate per route. That number, not the price list, tells you whether you got it right.

Those five rows are not everything we carry. The image side is grouped under text-to-image models, and every model we run, video included, is priced in the catalog. If you are weighing this against not paying at all, we wrote about what free AI image generators actually get you. If you have decided to pay and want to know which tier earns it, that is which AI image generators are worth paying for.

Frequently asked questions

What is the cheapest API for image generation?

The cheapest image generation request on our catalog is Nano Banana 2 Lite at $0.0238 per image, checked 26 August 2026. That covers both text-to-image and image editing, since editing carries no surcharge. It outputs at 1K only and has no resolution parameter. Prices move, so re-check before you build a budget on any figure.

How much does it cost to generate 1,000 AI images?

At $0.0238 per image, 1,000 images is $23.80. On Nano Banana 2 at $0.04 it is $40, and on Nano Banana Pro at $0.075 it is $75. Those figures assume one attempt per image, which is the assumption most budgets get wrong. Multiply by your real retry rate.

Why is GPT Image 2 priced differently than the base number suggests?

Because quality is a ×7 multiplier that applies at every tier, including low and auto. The base catalog figure of $0.0075 is never what anyone pays; a standard 1K request costs $0.0525. Resolution then multiplies on top, taking 2K to $0.105 and 4K to $0.1575.

Does 2K output cost extra on Nano Banana Pro?

No. Nano Banana Pro charges the same $0.075 for 1K and 2K, so requesting 2K is free additional resolution on that tier. 4K doubles it to $0.15. That is unusual. Nano Banana 2, by contrast, charges 1.5× for 2K and 2× for 4K.

Is the cheapest model always the cheapest option overall?

No, and the break-even is easy to compute. Four attempts on a $0.0238 model cost $0.0952, which is more than one $0.075 image from a model that gets it right first time. If a job needs in-image text, many references, or print resolution, the cheap tier fails every attempt rather than most of them, and the retries are where the money goes.

Does editing an image cost more than generating one?

No. On every image model we carry, the edit endpoint is priced identically to its text-to-image counterpart — $0.0238 on Nano Banana 2 Lite, $0.04 on Nano Banana 2, $0.075 on Nano Banana Pro. Passing reference images does not change the price of a request.

Does paying for a higher tier remove the SynthID watermark?

No. Every Google image model marks its output with SynthID, and the $0.075 tier is marked exactly like the $0.0238 one. It is not a setting anyone can switch off, here or at Google itself, so a provider advertising watermark-free Google output is selling something it cannot deliver. Settle that constraint with legal before you integrate rather than after.