New ListingRVA AI is now in Claude: generate listing copy and publish images by chat. Now in Claude. Set it up →
Lifetime Pay once, never pay for product photography again. Lifetime access on select plans. Pay once, not monthly. See lifetime →
← Blog
Data

Gemini vs OpenAI Image Generation: What We Learned From Product Photos

November 15, 2026 · 13 min read · by

Gemini vs OpenAI image generation cannot be settled on price per product photo from the official pages alone. Google lists its Gemini image models with a per-image equivalent (Gemini 3 Pro Image is $0.134 per 1K or 2K image), while OpenAI lists its image models per 1M tokens only, so any single per-image price you see quoted for OpenAI is somebody's estimate. What you can compare honestly is the unit each provider bills in, how fast each model returned images in real runs, and how often it got the product and the words on the image right.

The question usually arrives as a budget problem. A seller with a few hundred SKUs wants fresh white-background and lifestyle images for every listing, reads a comparison that names a winner, wires that model into a script or picks a tool built on it, and then the bill does not match the estimate, or the model is retired halfway through the catalogue and a weekend disappears into fixing it. This post puts both official price lists side by side with their units, as a snapshot taken on 4 November 2026, and adds what we saw running Gemini and OpenAI models in production, small samples flagged. For the wider money question of studio, freelancer or AI, start with the breakdown of product photography cost per SKU.

Key takeaways

  • Google lists Gemini image models with a per-image equivalent: Gemini 3 Pro Image is $0.134 per 1K/2K image and $0.24 per 4K image, and Gemini 3.1 Flash Image runs from $0.045 at 0.5K to $0.151 at 4K.
  • OpenAI lists its image models per 1M tokens only (gpt-image-1.5: $8.00 image input, $32.00 image output), so there is no official per-image number to set against Gemini.
  • Every price here is a 4 November 2026 snapshot. gpt-image-1.5 shuts down on 1 December 2026 and gemini-2.5-flash-image on 15 March 2027.
  • In our runs, Gemini Flash returned images fastest but misspelled a promo headline; Gemini Pro did not, and the GA Pro model scored highest in our first 14 runs.
  • A provider's API rate is not the price a seller pays inside a tool, so never multiply it into a plan or credit price.
Also ReadWhy 'Not as Pictured' Returns Keep Happening Even When Your Photos Are Accurate
›
Five image models and how each is priced: three Gemini models listed per image and two OpenAI models listed per 1M tokens.

How does Gemini vs OpenAI image generation pricing compare?

Gemini vs OpenAI image generation pricing compares at the level of units, not winners: both providers bill image output as tokens, but Google converts its token counts into a per-image equivalent on the pricing page and OpenAI does not. Google's page says a 2K image from Gemini 3.1 Flash Image consumes 1680 tokens and a 4K image from Gemini 3 Pro Image consumes 2000 tokens, and then states the dollar figure for one image. OpenAI's page lists gpt-image-1.5 and gpt-image-1-mini under the line "Prices per 1M tokens", with separate rates for image input, image output and cached input.

That difference matters more than it looks. A per-image number lets you budget a catalogue with one multiplication. A per-token number means the cost of each OpenAI image depends on how many tokens that image used, and OpenAI does not turn that into one official figure on the page. So this post does not give an OpenAI price per image, and it does not claim either provider is cheaper per image. If you need that number for your own catalogue, run a small batch at the size and settings you actually use, then multiply the tokens you were billed for by the listed rate.

It is also worth separating the model from the provider. Google sells several image models at very different prices, and so does OpenAI, so "Gemini or OpenAI" is really a choice between specific model IDs, each with its own price, speed, text handling and shutdown date. The rest of this post works at that level: the model ID, the unit it is billed in, and what happened when we ran it.

What does one Gemini image cost at 1K, 2K and 4K?

One Gemini image costs between $0.039 and $0.24 at list price, depending on the model and the output size, according to Google's pricing page as checked on 4 November 2026. These are the standard rates for each image that comes back:

ModelOutput sizeListed price (standard)
Gemini 3 Pro Image1K or 2K$0.134 per image
Gemini 3 Pro Image4K (up to 4096x4096px)$0.24 per image
Gemini 3 Pro ImageReference image input$0.0011 per input image
Gemini 3.1 Flash Image0.5K$0.045 per image
Gemini 3.1 Flash Image1K$0.067 per image
Gemini 3.1 Flash Image2K (2048x2048px)$0.101 per image
Gemini 3.1 Flash Image4K$0.151 per image
Gemini 2.5 Flash ImageStandard$0.039 per image (retiring 15 Mar 2027)

Two lines from the page are worth reading in Google's own words. For Gemini 3 Pro Image, "Image input is set at 560 tokens or $0.0011 per image.", so sending a reference photo of your real product adds very little to the bill. For large output, "Output images up to 4096x4096px (4K) consume 2000 tokens and are equivalent to $0.24 per image." The 1K and 2K rate is stated as "Equivalent to $0.134 per 1K/2K image". The full list, including the Flash models, sits on Google's Gemini API pricing page.

For budgeting, the arithmetic is simple. Ten 2K images on Gemini 3 Pro Image cost 10 times $0.134, which is $1.34, plus 10 times $0.0011 for the reference photos, about one cent more. The same ten images at 1K on Gemini 3.1 Flash Image cost $0.67. Those are list prices for images that come back. Every regeneration you run because a label came out wrong is billed again, so your real cost per usable image depends on how often the first attempt is good enough to publish.

Also ReadAI Image Artifacts: The Checklist to Run Before You Publish
›

Size is the other lever. Paying for 4K only makes sense if your marketplace or your zoom view needs it, and the jump from 2K to 4K on Gemini 3 Pro Image is $0.134 to $0.24 per image. Gemini 2.5 Flash Image is the cheapest line on the page at "$0.039 per image.", but it is retiring on 15 March 2027, so it is a poor base for anything you plan to run next year. Before you choose a size, check what your main image actually has to meet in the Amazon product image requirements.

How are OpenAI image models priced, and why is there no simple per-image number?

OpenAI prices gpt-image-1.5 and gpt-image-1-mini per 1M tokens, split into image input, image output and cached input, and its pricing page does not give a per-image equivalent for either model. These are the Standard tier rates from the snapshot:

Model (Standard tier)Image inputImage outputCached input
gpt-image-1.5 (shuts down 1 Dec 2026)$8.00 per 1M tokens$32.00 per 1M tokens$2.00 per 1M tokens
gpt-image-1-mini$2.50 per 1M tokens$8.00 per 1M tokens$0.25 per 1M tokens

Read the units carefully: $32.00 is the price of one million image output tokens on gpt-image-1.5, not the price of an image, and $8.00 is the same thing for gpt-image-1-mini. To turn either into a cost per product photo you need the number of tokens each of your images used, which only your own runs can tell you. The rates and the "Prices per 1M tokens." line are on OpenAI's API pricing page.

There are also two OpenAI models this post deliberately gives no price for. We ran gpt-image-1 in production until it shut down on 23 October 2026, and we tested gpt-image-2, but we did not confirm a price for either on the pricing page in this snapshot, so neither appears in the tables. And gpt-image-1.5, the model in the table above, shuts down on 1 December 2026, so a budget built on it has only weeks left.

The order ListingRVA AI uses image models: Gemini 3 Pro first, Gemini Flash second, OpenAI last, with credit refunds and shutdown alerts.

Which models do we run, and why Gemini Pro first, then Flash, then OpenAI?

ListingRVA AI runs paid image generation on Gemini 3 Pro Image, the free tier on Gemini 3.1 Flash Image, and Pro wherever on-image text matters, with OpenAI kept as the last fallback. Pro is first because it handled headline text without the spelling mistakes we saw from Flash. Flash is second because it was fast, and on Google's page it costs less per image than Pro at the same size. OpenAI is third because its image models kept moving under us: gpt-image-1, which we ran in production, shut down on 23 October, gpt-image-2 was dropped because it hung, and gpt-image-1.5 shuts down on 1 December.

The order is also a safety net. If a model fails, generation falls back from Gemini Pro to Gemini Flash and then to OpenAI, so the seller still gets an image, and if every attempt fails, the credits are refunded. Each model kind is a single config setting, so replacing a retiring model is a settings change rather than a rebuild, and the owner gets alerts ahead of announced model shutdowns.

If you call these APIs from your own script, copy the pattern rather than the model list. Keep the model ID in one place, set a timeout, name a second model for when the first fails, and decide what happens when both fail. Switching models also changes the look of the output, so a catalogue made half on one model and half on another can drift, and you need a plan to keep one catalogue consistent.

What happened when we ran the same photo through two models?

When we ran the same prompt and the same product photo through Gemini 3.1 Flash Image and Gemini 2.5 Flash Image on 29 October 2026, the two tied on white-background hero images, and both misspelled the promo headline. One rendered "ttalic" and the other "SSeriously", the kind of doubled or dropped letter that a shopper notices before they notice the product. Gemini 3 Pro did not make that mistake in our earlier runs.

Neither of those is an OpenAI model, and that limits what we can claim. Our OpenAI numbers come from production jobs on different products and prompts, not from a same-photo test against Gemini, so we can say how each model behaved for us, but we cannot say one provider wins on identical input. Any comparison that does claim that should show you the prompt, the photo and the output from both sides.

The practical reading for a seller is narrower and more useful. For a plain white-background main image with no words on it, the Flash models were good enough in that test, which is why the free tier runs on Flash. For anything with a headline, a price callout or a benefit line drawn into the image, Pro was the safer choice for us, and the cheaper per-image rate of Flash disappears quickly if you have to regenerate every image with text.

Where does each model fail on real products (text, labels, shape)?

In our runs, each model failed in its own way: Gemini Flash on spelling in headlines, gpt-image-2 by hanging, and gpt-image-1 by being retired. Any model can still get label text, logos or product shape wrong, so every image needs a check against the real product before it goes live:

  • On-image text: both Flash models misspelled a promo headline in our 29 October test. Use Pro for text, or add the words afterwards in an editor.
  • Hangs: gpt-image-2 had a median of 193.9 seconds across 17 images and was dropped because it hung. A script without a timeout simply waits.
  • Retirement: gpt-image-1 shut down on 23 October 2026. Requests to a retired model stop working, whatever your image needed.
  • Labels and fine print: read every word on the pack in the output against the real packaging, including ingredients, sizes and logos.
  • Shape and proportions: compare caps, handles, spouts and edges against your reference photo, since a wrong shape becomes a return.

None of these depend on the provider's price, and all of them cost more than the image itself when they reach a live listing: a misspelled headline makes a brand look careless, and a reshaped bottle invites "not as described" complaints. A short review routine catches most of them in a minute per image, and the full routine is in the AI image artifacts checklist.

Will these prices and models still exist next quarter?

Not all of them: gpt-image-1.5 shuts down on 1 December 2026, gemini-2.5-flash-image retires on 15 March 2027, and every price in this post is a snapshot from 4 November 2026. Image models now change faster than most sellers update a spreadsheet, and gpt-image-1 already shut down on 23 October 2026.

Shutdown dates to note: gpt-image-1.5 shuts down on 1 December 2026 and gemini-2.5-flash-image on 15 March 2027. If a script, plugin or tool you rely on names either model, plan the move now, and check the provider's page again on the day you set a budget.

A model change is not only a price change. When we moved from the preview version of Gemini 3 Pro Image to the GA model, both the speed and the quality numbers shifted, as the data section below shows. So when a model you depend on is replaced, rerun a handful of your own products through the new one, compare against your current images, and only then switch the whole catalogue.

Also ReadWhen to Reshoot Product Photos and When AI Can Rescue Them
›

A simple record helps: for every batch, note the model ID, the output size, the date and the listed price you budgeted with. When the next shutdown notice arrives, you will know exactly which images came from which model, which ones to regenerate first, and whether the new price changes your plan. The same record answers the question of whether a regenerated image still matches the rest of the listing.

What does this mean for the price a seller pays per image?

For a seller, the provider's API rate is what one returned image costs whoever calls the API, not the price you pay inside a tool, and it should never be multiplied into a plan or credit price. If you call the API yourself, your real cost is the list price for every image you generate, including the ones you throw away, plus the time spent on timeouts, retries and moving off retired models.

If you use a tool, you pay the tool's plan, and what that plan covers is up to the tool: which models it uses, what happens when one fails, and whether you are charged for failures. Compare plans on what you get each month and how failures are handled, not on the API rate underneath. ListingRVA AI refunds the credits for failed generations and shows what each plan includes on the page with current plans and credits.

Quick test: before you trust a per-image price for any OpenAI image model, ask where the number came from. If it does not come with a size, a token count and the listed rate, it is an estimate, not a price.

For many sellers the honest conclusion of the Gemini vs OpenAI image generation question is that the model matters less than the routine around it: the right size, a check on every image, a fallback, and a plan for shutdown dates. If you would rather not manage model IDs at all, it is worth taking time to compare the tools built on these models.

From our own data: how fast and how good was each model?

Between our launch in April and 30 October 2026, across the image jobs that recorded a timing, including our own testing, Gemini 2.5 Flash Image returned images fastest and the GA version of Gemini 3 Pro Image had the highest average quality score in its first 14 runs. The quality score comes from our automated image judge.

ModelImages with a timingMedian time90th percentileAverage quality score
gemini-3-pro-image-preview (the preview version we used before moving to the GA model)59125.7 s44.8 s86.6
gemini-3-pro-image (GA, October)First 14 runs67.5 s79.6 s97.9
gemini-2.5-flash-image999.9 s21.1 s85.9
gpt-image-1 (retired 23 Oct)1153.6 sNot reported94.5
gpt-image-2 (dropped because it hung)17193.9 sNot reportedNot reported

Read the small rows as early signals. In our first 14 runs of the GA Pro model, the median time was 67.5 seconds against 25.7 seconds for the preview version across 591 images, while the average quality score rose from 86.6 to 97.9. gpt-image-1 ran on 11 images before it was retired and gpt-image-2 on 17 before we dropped it. Each row covers different products and prompts, so this is what we saw, not a benchmark.

What we changed as a result is the setup described above: paid images on Gemini 3 Pro, the free tier on Gemini 3.1 Flash Image, Pro wherever on-image text matters, and OpenAI as the last fallback.

Source: ListingRVA AI generation records (totals only, no customer data).

Get product images without managing models

Generate marketplace-sized white background, lifestyle and infographic images, A+ pages and listing copy, with fallbacks and refunds for failed generations handled for you. Start free with 50 credits, no credit card.

Start free →

Frequently asked questions

In Gemini vs OpenAI image generation, which costs less per product photo?

Neither can be named from the official pages. Google lists per-image equivalents, such as $0.134 per 1K/2K image for Gemini 3 Pro Image, while OpenAI lists gpt-image-1.5 and gpt-image-1-mini per 1M tokens only. Without the token count of each OpenAI image there is no like-for-like per-image comparison, so run your own batch and multiply the tokens billed by the listed rate.

How much does a Gemini 3 Pro Image cost at 4K?

Google's pricing page says 4K output images "consume 2000 tokens and are equivalent to $0.24 per image." At 1K or 2K the rate is $0.134 per image, and a reference image input is $0.0011 per image. These figures are a snapshot from 4 November 2026.

What are OpenAI's image model prices?

On the Standard tier, per 1M tokens: gpt-image-1.5 is $8.00 for image input, $32.00 for image output and $2.00 for cached input; gpt-image-1-mini is $2.50, $8.00 and $0.25. OpenAI does not list a per-image equivalent. gpt-image-1.5 shuts down on 1 December 2026.

Which model handled text on product images better in your tests?

Gemini 3 Pro. In our 29 October 2026 test with the same prompt and photo, Gemini 3.1 Flash Image and Gemini 2.5 Flash Image tied on white-background hero images, but both misspelled the promo headline. Pro did not in our earlier runs, so we use Pro wherever on-image text matters.

When do gpt-image-1.5 and gemini-2.5-flash-image shut down?

gpt-image-1.5 shuts down on 1 December 2026 and gemini-2.5-flash-image on 15 March 2027. gpt-image-1 already shut down on 23 October 2026. Check each provider's page again before you budget, because these dates and prices can change.

How fast is Gemini 3 Pro Image?

In our first 14 runs of the GA model in October 2026, the median time was 67.5 seconds and the 90th percentile 79.6 seconds. The preview version we used before moving to the GA model had a median of 25.7 seconds across 591 images. The GA sample is small, so treat it as an early signal.

Can I turn these API prices into ListingRVA AI credit costs?

No. API prices are what a provider charges whoever calls its API, not what a seller pays in a tool. ListingRVA AI plans and credits are listed on its pricing page, and credits for failed generations are refunded.

Keep going

Browse all articles →
New

ListingRVA AI now works inside Claude. Connect your workspace and ask Claude to generate listing copy and publish product images to your store, without leaving your editor. See the Claude connector →

Share

Comments

No comments yet, be the first.

Leave a comment

Comments are reviewed before they appear.

Hi! Questions about product photography or your listings? Ask AI anything.