← Blog

Best way seriesNovember 16, 2026

The best way to choose an AI image model

Default to nano-banana-pro at 33¢ an image, switch to gpt-image-2-5 Flare at 10¢ for readable text, and use a 4¢ model only when its output meets the brief.

For website, ad and product images, have your agent call fal/generate with nano-banana-pro (33¢ per image) by default, and switch to gpt-image-2-5--flare--text-to-image (10¢) when the image has to carry readable text. A cheaper model is enough when its output already meets the brief, and the catalog lists one at 4¢ for that case.

Those picks come from the routing guidance that Vaaya's consult tool works from. Vaaya is a wallet for AI agents: you recharge it with a prepaid pack and your agent pays per call from one balance. There is no subscription and no separate account or API key for each model provider.

What "best" means for an image

  • Fit to the job. A photoreal scene and a graphic with readable words are different jobs, and the guidance routes each to a different model.
  • The size you need. Some models take an aspect ratio and a resolution tier. Others take pixel dimensions.
  • Cost per accepted image. Count the revision. An edit is a separate call at the edit model's price.
  • Consistency across a set. A page reads as one brand when its images share a palette and a style.

The image models in the catalog

Image generation on Vaaya goes through one action, fal/generate. The model parameter picks the model and sets the price. Prices here are as listed on 16 November 2026. The catalog has current prices, and the fal/generate page lists every model.

fal/generate with model What it returns Price per call When to pick it
nano-banana-pro A photoreal image from a prompt, up to 4K 33¢ The default for website and marketing images: heroes, backgrounds, people shots, section art
nano-banana-pro--edit An edited version of the images in image_urls 33¢ The default edit: revisions, and a series that must keep one character or style
gpt-image-2-5--flare--text-to-image An image with readable text in it 10¢ New work that needs words in the image, such as labels, diagrams or a headline on a social card
gpt-image-2-5--flare--edit An edited image, text and labels included 11¢ Adding or correcting text on an image you already have
gpt-image-2 An image with readable text in it 24¢ Readable text, as the earlier text model. The guidance prefers 2.5 Flare for new work
seedream--v5-pro--text-to-image A photoreal image, up to 2K 14¢ Realistic people and scenes, and a still you plan to animate into video
seedream--v5-pro--edit An edit or composite of the images in image_urls 14¢ Editing a photo while keeping it realistic
seedream--v4-5--text-to-image An image from a prompt 4¢ Bulk work and iteration, when its output already meets the brief

Each call returns the image URL in the same response. Two more entries help with website work. gpt-image-2-5--sunburst--text-to-image (10¢) is the higher-detail variant of the 2.5 model. image-background-removal (5¢) cuts a subject out of an image and returns a transparent PNG.

The default path

Start with nano-banana-pro. The guidance treats it as the default for nearly every website and marketing image. You size it with aspect_ratio and resolution (1K, 2K or 4K). The price is 33¢ at every tier, so generate at 4K and downscale when you save the file.

Change models when the job changes:

  • Words inside the image: gpt-image-2-5--flare--text-to-image.
  • A realistic person or scene that becomes the first frame of a video: seedream--v5-pro--text-to-image.

Price comes last. The guidance is to choose on what the image has to do, and to take the cheaper option only between models that both meet the brief. That is the case for seedream--v4-5--text-to-image at 4¢ when you are producing many images or iterating on a direction.

Sometimes the cheapest image is no model call at all. The guidance says to write flat gradients, solid fills and standard UI icons as CSS or SVG, and to spend generation on imagery that needs a model.

A worked example: four images, one revision each

This is arithmetic at listed prices on 16 November 2026. It reports no measured results.

A landing page needs a hero photo, two section illustrations and a social card with a headline.

First pass:

  • Hero: nano-banana-pro, 16:9 at 4K: 33¢
  • Two section illustrations: nano-banana-pro, 2 × 33¢ = 66¢
  • Social card: gpt-image-2-5--flare--text-to-image: 10¢

One revision each:

  • Hero and both illustrations: nano-banana-pro--edit, 3 × 33¢ = 99¢
  • Social card: gpt-image-2-5--flare--edit: 11¢

Total: 109¢ + 110¢ = 219¢, or $2.19.

The hero call:

{
  "service": "fal",
  "action": "generate",
  "params": {
    "model": "nano-banana-pro",
    "prompt": "Wide photo of a barista pouring milk into a cup at a wooden counter, soft morning window light, 35mm lens, shallow depth of field, clear wall on the left for a headline",
    "aspect_ratio": "16:9",
    "resolution": "4K"
  },
  "max_cost_cents": 33
}

The revision passes the image URL from the first response:

{
  "service": "fal",
  "action": "generate",
  "params": {
    "model": "nano-banana-pro--edit",
    "prompt": "Same photo. Make the cup plain white with no logo. Change nothing else.",
    "image_urls": ["<image URL from the first response>"]
  },
  "max_cost_cents": 33
}

Suppose the two illustrations already meet the brief on seedream--v4-5--text-to-image, with revisions on seedream--v4-5--edit (4¢). Those four calls then cost 16¢ in place of 132¢, and the set comes to $1.03.

An image from an earlier call can go straight into an edit. A file from your own disk, such as a logo or a product photo, is staged first with fal/upload (1¢). Upload the original, uncompressed.

How to run it

Connect an agent over MCP, the API or the SDK. Then ask in plain English:

Use Vaaya to make four images for my landing page: a 16:9 hero photo, two section illustrations in one style, and a 1200 by 630 social card with the headline "Fresh roasts, shipped weekly". Take the brand colors from my repo. Allow one revision per image. Tell me the model and the charge for every call.

The agent can ask Vaaya's consult tool to plan the calls, and the plan comes back with a price on each one. Every use call carries max_cost_cents. A call priced above that number is refused before it runs and costs nothing. For a slower walk through one image and one revision, see how to generate and revise an image.

What to check, and where this falls short

  • Save the file at once. The image URL arrives in the response. Calling again to recover a lost URL is a new paid generation.
  • Exact sizes. nano-banana-pro takes ratios and tiers. Generate the closest ratio at 4K, then crop and downscale. A 1200 by 630 card starts as 16:9. Extreme banners such as 728 by 90 cannot be generated directly.
  • Text. Read every word in the output. For a precise diagram, an exact wordmark or a real data chart, the guidance is to write SVG.
  • gpt-image-2 is slow. Run those calls one at a time. A single call can time out.
  • Refusals. A content_policy_violation on nano-banana-pro charges nothing. Reword the prompt and retry.
  • Real things. Use a screenshot for your actual product UI, and real photos for real, named people.
  • The set. View the images together and regenerate any outlier.

Video is a separate choice. The video model comparison covers it.

Why this job suits a pay-per-call wallet

Image work comes in bursts: a launch week, then nothing for a month. The right model also changes from one image to the next, so a single balance that reaches all of them fits the job better than a plan with one image tool. Revision loops are where spend creeps. A daily limit on the agent's key and max_cost_cents of 33 on each call put a hard edge on that, as spending controls explains. Packs and per-call billing are on the pricing page.

Questions

Which AI image model should an agent use for website images?

Vaaya's routing guidance makes nano-banana-pro the default for website and marketing images, listed at 33¢ per image on 16 November 2026. When the image has to carry readable text, it points to gpt-image-2-5--flare--text-to-image at 10¢.

How much does one AI-generated image cost?

It depends on the model you pass to fal/generate. At prices listed on 16 November 2026, the main image models cost between 4¢ (seedream--v4-5--text-to-image) and 33¢ (nano-banana-pro) per image. A revision is a separate call at the edit model's price.

When is a cheaper image model enough?

When its output already meets the brief. The guidance is to choose on what the image must do first and to let price decide only between models that both do the job. The 4¢ seedream--v4-5--text-to-image is the listed pick for bulk work and iteration.

Can an image model produce exact pixel sizes?

nano-banana-pro takes an aspect ratio and a resolution tier in place of pixel dimensions. Generate the closest ratio at 4K, then crop and downscale to the target. Extreme banner shapes such as 728 by 90 cannot be generated directly.

One balance for every image model

Connect an agent, put a limit on its key, and let it pay per image for the model each job needs.

Connect an agent