AI Image Generation: 6 Models Compared — From 500 to 15,000 Credits

TL;DR: 6 AI image models on one credit balance, from 500 to 15,000 credits. Start on Nano Banana 2 Lite to find your prompt, finish on Nano Banana Pro (the only 4K upscale) or GPT Image 2 (best text rendering).

The line-up at a glance

Every price is flat per image. Failures are refunded automatically.

Nano Banana 2 Lite — 500 credits

The cheapest image on the platform and the fastest (15-30s). Use it to iterate: run a prompt five times here for the price of one GPT Image 2 render, then re-run the winner on a better model.

Nano Banana 2 — 1,000 credits

The general-purpose default. ~1K native output (1376x768), 30-60s, four reference images. If you are not sure which model to use, this is the one.

Grok Image — 1,000 credits

xAI's image model, same price as Nano Banana 2 but with 2:3 and 3:2 natively — the two ratios the Nano Banana family cannot draw. Attach a reference and it switches to image-to-image automatically.

Meta AI — 1,500 credits

Meta handles references differently from everything else: instead of a flat list, you pass named components — character_image, scene_image, style_image — each binding to a specific role in the composition. That is far more controllable than "here are four pictures, good luck". Square, portrait and landscape only.

Nano Banana Pro — 3,500 credits

The quality pick in the Nano Banana family, and the only model with a 4K upscale (resolution: "4K"). Slower at 60-90s and more faithful to long, detailed prompts. Recently reduced from 5,000 credits.

GPT Image 2 — 15,000 credits

The premium option, and the one to reach for when the image contains text — signage, packaging, UI mockups, posters. It is also the only model accepting all seven aspect ratios.

Output is capped at ~1.57 MP and follows the aspect ratio; there is no 2K/4K tier and no upscale, so pick Nano Banana Pro if you need 4K.

Four controls are unique to it:

• quality — low / medium / high (default high). Changes the output, not the price.
• prompt_mode — auto lets the model rewrite your prompt for better results; direct uses it verbatim. Use direct when you have already tuned the wording.
• reasoning — none through max. Higher settings follow complex prompts more closely but take longer.
• web_search — grounds the generation with a web search first, useful when referencing something real and current.

Recently reduced from a tiered 4,500-80,000 matrix to a flat 15,000.

Which should you pick?

• Iterating on a prompt? Nano Banana 2 Lite at 500.
• No strong opinion? Nano Banana 2.
• Need 2:3 or 3:2? Grok Image.
• Controlling character vs scene vs style separately? Meta AI.
• Need 4K? Nano Banana Pro — nothing else upscales.
• Image contains readable text? GPT Image 2.

The cheap-first workflow

The thirty-fold price gap between the cheapest and most expensive model here is not a ranking, it is a workflow.

A prompt almost never works first time. The composition is off, the subject is facing the wrong way, the mood is not what you pictured. Finding that out on a 15,000-credit model costs thirty times what finding it out on a 500-credit one does, and you learn exactly the same thing.

So iterate on Nano Banana 2 Lite. Run five variations, decide what the image actually needs to be, then render the winner once on whichever model suits it. Five drafts plus one premium render costs less than two premium renders, and the result is better because you knew what you were asking for by the time you asked.

Reference images do more than style transfer

Every model here except Meta takes a flat list of reference images, and the common mistake is treating them as a style dial. They are closer to a specification.

Give it the product and it keeps the product's actual shape and label rather than inventing a plausible one. Give it a face and it keeps the same person across a set. Give it a room and the next image is that room from a different angle. This is what makes a consistent set possible at all, and prompt text alone cannot do it.

Meta is the exception worth knowing: it binds references to named roles — character, scene, style — rather than a list. That is more work to set up and much more controllable when you are recombining a fixed cast against changing backdrops.

Aspect ratio is chosen for you more than you think

Not every model renders every ratio, and asking for one a model cannot do gets substituted rather than rendered.

The nano family covers the everyday ratios but not the wide cinematic ones. Grok is the one to reach for when you need 2:3, 3:2 or ultra-wide. Meta is square and the two standard widescreens only. GPT Image 2 covers the widest set.

Decide the ratio before the model, not after — it is the constraint that eliminates most of the list.

For the moving-image equivalent of this decision, see [the AI video model comparison](/blog/ai-video-generation-veo3-grok-sora2-seedance-complete-guide).

Get Started

All 6 models are in the [image dashboard](/dashboard/image) and the [public API](/dashboard/api-docs), on one credit balance with no per-model subscription. A practical workflow: draft on Lite, pick the winner, re-render on Pro or GPT Image 2.