A hundred images costs you 300 credits on Nano Banana 2 Lite, or 400 credits on GPT Image 2.5 in Draft, or 600 credits on GPT Image 2.5 in Standard. Same prompt box, same site, same afternoon. The gap between the cheapest and the most expensive of those three is exactly double.
Put a currency on it. On the Standard plan — $29.90 for 300 credits a month — a credit works out to just under $0.10. So a hundred Lite images is $29.90 and eats your entire monthly allowance. A hundred Draft images is $39.87. A hundred Standard-tier GPT Image 2.5 images is $59.80. And a hundred at Ultra, the top tier, is 2,200 credits, or roughly $219. That last number is why this article exists: at volume, the tier you pick matters far more than the model you pick.
This is written on September 10, 2026, two days after OpenAI shipped GPT Image 2.5, and with Nano Banana 2 Lite generally available from Google as the Gemini 3.1 Flash-Lite Image model. Both are selectable here, so every credit figure below is checkable in about a minute. What follows is not a quality shootout. It is a spreadsheet: what a dollar buys, what you give up when you spend less, and the three workflows where spending less costs you nothing.
The price sheet
Here is the whole thing on this site, in credits per image. Image-to-image on GPT Image 2.5 costs the same as text-to-image at every tier.
| Model / tier |
1K |
2K |
4K |
| Nano Banana 2 Lite |
3 |
— |
— |
| GPT Image 2.5 — Draft |
4 |
5 |
6 |
| GPT Image 2.5 — Standard |
6 |
8 |
14 |
| GPT Image 2.5 — Ultra |
22 |
24 |
42 |
| Nano Banana 2 (full) |
6 |
— |
— |
Lite has one price. Three credits, any scene, no tier menu to get wrong. Google generates it at a fixed 1K, so there is no resolution column to shop.
Now the same table as cost per 100 images, in credits, with the dollar equivalent at the Standard plan rate of about $0.0997 per credit:
| Model / tier |
100 images (credits) |
≈ at plan rate |
| Nano Banana 2 Lite |
300 |
$29.90 |
| GPT Image 2.5 Draft 1K |
400 |
$39.87 |
| GPT Image 2.5 Draft 2K |
500 |
$49.83 |
| GPT Image 2.5 Draft 4K |
600 |
$59.80 |
| GPT Image 2.5 Standard 1K |
600 |
$59.80 |
| GPT Image 2.5 Standard 2K |
800 |
$79.73 |
| GPT Image 2.5 Standard 4K |
1,400 |
$139.53 |
| GPT Image 2.5 Ultra 1K |
2,200 |
$219.27 |
| GPT Image 2.5 Ultra 4K |
4,200 |
$418.60 |
Three things fall out of that table that are easy to miss.
Lite is exactly half of Standard 1K. Three credits against six. It is also exactly half of full Nano Banana 2 at 1K, which is likewise 6 credits. Whenever you choose Lite over either of them, you are buying two images for the price of one, and the entire question is whether the second one is worth having.
Draft is barely cheaper than Standard, but Ultra is enormously more expensive than both. Draft 1K to Standard 1K is 4 credits to 6 — a 50% step. Standard 1K to Ultra 1K is 6 to 22 — a 3.67× step. The interesting budget decision on GPT Image 2.5 is not Draft versus Standard. It is whether you touch Ultra at all.
Resolution is nearly free in Draft and brutally expensive in Ultra. Draft 1K is 4 credits and Draft 4K is 6. Two extra credits buys you sixteen times the pixels. Draft 4K also costs the same 6 credits as Standard 1K. In Ultra, the same jump runs 22 credits to 42. If you need big files on a budget, Draft is the lane where big files are cheap.
What the small purchases actually buy
Credit packs run $9.90 for 80 credits, which is about $0.124 per credit — a bit more than the plan rate, because you are not committing to a subscription. That pack buys:
- 26 Nano Banana 2 Lite images (78 credits, 2 left over)
- 20 GPT Image 2.5 Draft 1K images (exactly 80)
- 16 Draft 2K images
- 13 Standard 1K images
- 10 Standard 2K images
- 3 Ultra 1K images (66 credits, 14 left over)
- 1 Ultra 4K image, with 38 credits stranded
Three Ultra renders for ten dollars is the number to keep in your head. It is not a bad deal for a finished piece. It is a terrible deal for exploration.
The free weekly check-in gives 30 credits. That is 10 Lite images, or 7 Draft 1K images with 2 credits left, or 5 Standard 1K images, or 1 Ultra 1K image. Claimed every week for a year, that is 1,560 credits — 520 Lite images or 390 Draft 1K images annually, at no cost. For a student or a side project, that is a real content budget, and it only works if you stay in the cheap lane. Full plan and pack details are on the pricing page.
What Draft mode and Lite actually are
The two cheap options got cheap in completely different ways, and confusing them is how people end up disappointed.
GPT Image 2.5 Draft is the low quality setting on a flagship. OpenAI ships GPT Image 2.5 with quality tiers — auto, low, medium, high, xhigh, max — and the tier is a compute dial, not a different model. In OpenAI's own API pricing, a 1024×1024 image runs roughly $0.006 at the low tier, about $0.053 at high, and about $0.21 at max. That is a 35× spread across one model. You are buying fewer output tokens: less time spent resolving detail, not a smaller or dumber network. Draft on this site maps to that cheap end of the same two engines OpenAI shipped on September 8, 2026, gpt-image-2.5-flare for speed and gpt-image-2.5-sunburst for precision.
Nano Banana 2 Lite is a purpose-built cheap model. It is Google's Gemini 3.1 Flash-Lite Image, the efficiency tier of the Nano Banana family, and it was designed from the start for high-volume generation rather than throttled down from something larger. Google says it generates in as little as four seconds, outputs at a fixed 1K across 14 aspect ratios, and carries SynthID watermarking by default. Google's stated strengths for it are world knowledge for quick contextual scenes, character consistency across rapid batches, and text that is legible — the word Google uses — for things like localized ad variations. It handles reference images for editing and style transfer, and supports fast multi-turn local edits.
The philosophies differ in a way that shows up in practice. Draft is a flagship holding back. Lite is a small model going flat out. That means Draft output tends to have the flagship's compositional instincts with less finish, while Lite output tends to be internally consistent and complete but built to a lower ceiling. Neither is a defect. They are different products with the same price tag.
Selection lives in different places too. GPT Image 2.5's tiers are on its own page at GPT Image 2.5. Lite does not have a separate tool page — it is a model option you pick on the Nano Banana 2 page.
Where cheap breaks down
Cheap generation is not free of consequence. Here is where you should expect to see the savings, stated as what the tiers are designed for rather than as test results.
Small text and dense typography. This is the sharpest edge. GPT Image 2.5's headline claim from OpenAI is sharper typography and small print — but that claim describes the model doing its best work, not the low tier skimming. On the Google side, Google's own language for Lite is that it renders legible text for rapid variants, with finer typographic control positioned as the job of the larger Nano Banana models. Hands-on comparisons of the Lite tier against full Nano Banana 2 and Nano Banana Pro consistently report that Lite has more trouble with small lettering and intricate logos. If your image has a price, a URL, a legal line, or a logo in it, the cheap lane is where you will get burned.
Fine detail at 100% zoom. Lower tiers spend fewer tokens resolving texture. Fabric weave, hair, foliage, reflections in glass, the edges of overlapping objects — these are the places rendering budget shows up first. At thumbnail size on a feed, most of it is invisible. At full-bleed print size, it is not.
Faces and hands. The standard advice applies and it is a budget question, not a model-quality question: any tier can produce a good face, but the cheaper the tier, the more attempts it takes to get one, and the more attempts you need, the less you actually saved. If faces are the point of the image, count your retries before you count your per-image price.
What does not break. Composition, colour, mood, layout, and whether the idea reads at a glance — the cheap tiers are perfectly capable at all of it. That is precisely why the workflows below work.
The volume workflows
Three patterns where the arithmetic is decisive rather than a matter of taste.
1. The social calendar workflow
Thirty days of posts, three concepts each. Ninety images.
Running all ninety on GPT Image 2.5 Standard 1K costs 540 credits — 1.8× the entire Standard plan monthly grant of 300. You would be buying packs before the month ended.
Running all ninety on Nano Banana 2 Lite costs 270 credits. Then take the five posts that actually matter — the launch, the offer, the two you are boosting, the profile banner — and re-render those at GPT Image 2.5 Standard 1K for 6 credits each, another 30 credits.
Total: 300 credits. That is exactly one month of the Standard plan, $29.90, with nothing left over and nothing bought on top. Against the all-Standard version, you saved 240 credits, or 44%, and the only images that took the cheaper path are the eighty-five nobody was going to zoom into.
2. Thumbnail iteration
Thumbnails are an iteration problem. You do not know which composition wins until you see twelve of them side by side, and eleven of the twelve get thrown away.
Twelve composition variants on Lite: 36 credits. Pick the winner, re-render it with real typography on GPT Image 2.5 Standard 2K at 8 credits: 44 credits total, about $5.45 at pack rate.
Twelve variants directly on Standard 2K: 96 credits, about $11.88. Twelve on Ultra 2K: 288 credits, about $35.64.
So the cheap-first path costs 54% less than all-Standard and 85% less than all-Ultra. The reason it works here specifically is that thumbnail text is exactly the thing Lite is weakest at — so you never ask Lite for the text. You ask it for the framing, the subject placement, the colour block, and then you pay once for the version with words on it.
3. Draft ten, upgrade one
The general form of the pattern, and the one worth internalising.
Generate ten concepts on Lite at 3 credits: 30 credits. Re-render the winner on GPT Image 2.5 Ultra 1K at 22 credits. Total 52 credits, about $5.18 at plan rate.
Generate all ten on Ultra: 220 credits, about $21.93.
That is 168 credits saved, a 76% reduction, for one finished image of identical quality — because the finished image came out of Ultra either way. The exploration is what you moved downmarket.
Here is the break-even that makes it obvious. The Lite drafting pass costs 30 credits, which is 1.36 Ultra renders. If drafting cheaply stops you from wasting even two Ultra renders on ideas that were never going to work, it has already paid for itself. In practice it stops you from wasting eight.
One caveat on this workflow: the concept you approve on Lite is not always the concept Ultra will hand back, because you are switching models, not just switching quality. If exact continuity matters, run the drafting pass on GPT Image 2.5 Draft at 4 credits instead — 40 credits for ten, 10 more than Lite — and stay inside one model family the whole way. That is what the extra credit per image is for.
When to spend more
Short section, because the honest answer is short.
Spend more when the image is the deliverable rather than a step toward it. Client work, a paid ad, a book cover, a print piece, a store listing hero — anything a stranger will judge you by, and anything you cannot cheaply redo.
Spend more when there is text in the image that has to be correct. Prices, dates, URLs, product names, legal copy. This is the single most reliable reason to leave the cheap lane, and it is the one place where paying 22 credits instead of 3 is not extravagance.
Spend more when you need genuine resolution for print or large-format display, and you need the detail to survive the zoom — though note that Draft 4K at 6 credits exists precisely so that "big" and "expensive" do not have to travel together.
Spend more when the retry count is climbing. If you are on your fifth Lite attempt at 3 credits, you have spent 15 credits and have nothing. Two Standard attempts would have cost 12. Cheap stops being cheap the moment you stop getting it in one or two passes, and the tell is always the same: you keep re-prompting the same thing.
Everything else — mood boards, storyboards, layout tests, A/B variants, blog headers, internal decks, Discord posts, anything you will look at once — belongs in the cheap lane. There are more than a dozen models here, and the skill worth having is knowing which ones you can afford to be careless with.
FAQ
What is the cheapest AI image generator on this site?
Nano Banana 2 Lite, at 3 credits flat per image regardless of scene complexity. The next cheapest is GPT Image 2.5 in Draft at 4 credits for 1K. At the Standard plan rate of about $0.0997 per credit, that is roughly $0.30 and $0.40 per image respectively, or 300 and 400 credits per hundred images.
Is GPT Image 2.5 Draft the same thing as Nano Banana 2 Lite?
No. Draft is the low quality setting on OpenAI's flagship — the same model, given less compute, with the tier acting as a dial. Lite is a separate model, Google's Gemini 3.1 Flash-Lite Image, built specifically for fast high-volume work. Same rough price, different reason for being cheap.
How many images do I get from the free weekly credits?
The weekly check-in gives 30 credits, which is 10 Nano Banana 2 Lite images, 7 GPT Image 2.5 Draft 1K images, 5 Standard 1K images, or 1 Ultra 1K image. Claimed every week for a year, that is 1,560 credits — 520 Lite images or 390 Draft images.
Is Nano Banana 2 Lite good enough for social media content?
For most of it, yes. Google positions it for rapid ideation, A/B testing and localized ad variants, and it generates in as little as four seconds at 1K across 14 aspect ratios. The exception is anything with small text or a detailed logo, where hands-on comparisons find Lite weaker than the larger Nano Banana models — put those through GPT Image 2.5 instead.
Can I use cheap-tier images commercially?
Commercial licensing on this site is tied to the plan, not the model or tier — it comes with Standard and above, and applies to Lite and Draft output the same as to Ultra. Google also embeds an invisible SynthID watermark in Nano Banana output by default, which identifies it as AI-generated without affecting how you can use it. Check the pricing page for current plan terms.
Run the same prompt through both and price it yourself: GPT Image 2.5 for the tier ladder, Nano Banana 2 for Lite. Ten Lite drafts cost 30 credits, and the weekly free credits cover that.
Sources
- Nano Banana 2 Lite and Gemini Omni Flash available — Google Cloud Blog — Google's GA announcement: four-second generation, SynthID by default, and the stated capabilities (world knowledge, character consistency, legible text and localization) plus target use cases.
- Nano Banana 2 Lite (Text to Image) on fal — fixed 1K (1024×1024) output, 14 supported aspect ratios, sub-2-second latency positioning, and the model's framing as the efficiency tier of the family.
- Gemini 3.1 Flash-Lite Image — Google DeepMind — the official model page confirming Nano Banana 2 Lite is Gemini 3.1 Flash-Lite Image.
- Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) — OpenRouter — third-party API token pricing for the Lite model, used as the outside reference point for its cost tier.
- ChatGPT Images 2.5: Faster, more precise, but not the same for everyone — The Decoder — GPT Image 2.5's quality tiers including the new xhigh and max, and the approximate per-image API costs at 1024×1024 of $0.006 low, $0.053 high and $0.21 max.
