You don’t need the top AI image model

• Only 4 of 27 scored models are worth their price. Every other model has something cheaper that scores higher.
• Gemini Nano Banana 2.1 is 46 Elo behind the top model for 1/6 of the price. OpenAI's premium buys first place, not a different class of image.
• The best open-weight model is 178 Elo behind. Open image models trail by about a year, and most can't be used commercially.
• Most models refuse a woman in a bikini. Only Gemini Nano Banana 2.1 drew all eight of our borderline prompts.
• Lettering is solved. All four models we sampled spelled a logo right on the first try.
• Editing photos of people is where they differ. Three models kept every face; three didn't.

35 models, 4 worth the price

Prices run from $0.007 to $0.28 per image, a 40x spread. Quality comes from the Artificial Analysis image arena, where people pick the better of two images made from the same prompt and the votes become an Elo score. 27 of the 35 models have one. Put the two together and most of the field falls off the best-value line.

View the in-depth comparison

• Recraft V4.1 Flash, $0.007, Elo 950. The cheapest image on the gateway, made in about two seconds. Fine for drafts and thumbnails.
• Muse Image 1.0, $0.01, Elo 1127. A one-cent image that scores within 47 Elo of Gemini Nano Banana 2.1, at a third of its price.
• Gemini Nano Banana 2.1, $0.034, Elo 1174. The sweet spot. Nothing cheaper scores higher, and the cheapest model that does costs 6x as much.
• GPT Image 2.5 Sunburst, $0.21, Elo 1220. The best image model by arena vote, in both text-to-image and editing.

Same prompt, four models: the differences are taste, not mistakes

We sent three prompts to the four models we made the defaults and alternatives on ModelCompare. All four spelled “ORBIT ROASTERS” correctly, which used to be the classic way to catch an image model out. All four made a convincing photo. What changes from model to model is style: warm and cinematic, or dark and moody.

Editing a photo of people is the real test

Most image models can now edit a picture you upload. The edit that tells them apart is one where people have to stay the same people. We gave six models a photo of three friends at a café and asked them to move the scene.

GPT Image 1 Mini is one of the cheapest editors on the list, which makes it tempting, but it redraws the people rather than keeping them. Use it for objects and scenes, not for photos of anyone who will notice their face changed.

A scene edit with no people in it is easier. We asked three models to turn Muse Image 1.0's rainy bookshop into a snowy morning, and all three kept the shop, the signs and the camera angle, and added believable snow.

Open-weight image models are a year behind

With text models, the best open-weight models have nearly caught up with the closed ones. With images they haven't. The best open-weight model in the arena, Qwen-Image-2.1, scores 1042: 178 Elo behind the top and 85 behind the one-cent Muse Image 1.0. That puts it a little above GPT Image 1, which came out in April 2025.

Most of these come with a licence that rules out commercial use: of the top seven, only Z-Image Turbo and the original Qwen Image are Apache 2.0, and they are the two weakest. Running one yourself means a GPU instead of a per-image bill, which can be cheaper at volume.

Three small open-weight models from Black Forest Labs are also on the gateway, so we ran them on the same prompts and edits as the closed models above, for $0.20 in total. They are the fastest and among the cheapest images we made: 3 to 8 seconds and one or two cents each.

The photos and the product shot hold up well next to the closed models. Lettering is where size shows: FLUX 3 Image Klein spelled “Orbit Roasters” cleanly, the 9B model mixed upper and lower case (“ORbIT”), and the 4B model wrote “ORDIT”. So lettering is solved for the big models, not for small ones you can run on your own machine. All three kept the three friends recognisable on the beach, though the two FLUX.2 models put their faces in the sunset's shadow, as FLUX Kontext did.

How strict are they? Most refuse a bikini

Every lab filters what its model will draw, and the filters are much stricter than most people expect. We sent ten models eight prompts that sit near the line without crossing it: a shirtless man at the gym, a 1950s pin-up, a couple kissing, a woman in a bikini, a Venus painting, a mother breastfeeding, a lingerie ad and a classical nude statue. Nothing explicit, and all of it the kind of picture you see in a magazine, a museum or an ad.

Gemini Nano Banana 2.1 drew all eight, and drew them as asked: no swapped outfits or cropped figures. Every other model refused at least two. MAI-Image-2.6-Flash drew only the shirtless man. OpenAI's models, both the top-ranked GPT Image 2.5 Sunburst and GPT Image 2, refused the bikini, the kiss and the breastfeeding mother. The shirtless man was drawn by nine of ten models, while the woman in a bikini was refused by eight of ten.

Two results go against reputation. Grok is marketed as the least filtered AI, yet Grok Imagine Image 2.0 drew two of eight. It also makes the picture first and then throws it away: its refusals took about 14 seconds, the same as a finished image, while ByteDance's refused in about a second by checking the prompt. And Seedream drew the lingerie ad but refused the shirtless man.

The open-weight FLUX models refused just as often as the closed ones, because we ran them on Black Forest Labs' own servers with its filter. Download the same weights and run them on your own machine and there is no filter at all. That is one reason people run image models themselves.

This is one run of each prompt through the API with default settings, for $1.36 in total. Filters that check the finished image can give a different answer on a second try, and apps like ChatGPT and Gemini add their own rules on top, so read it as a picture of how strict each lab is, not a fixed rule.

From two seconds to two minutes

Arena scores don't tell you how long you wait. We timed real runs over two rounds of testing, one 1024×1024 image at a time.

The two slowest are not the two priciest. Seedream 5.0 Pro costs $0.035 and took anywhere from 40 seconds to over two minutes. GPT Image 2 at high quality takes about a minute and a half every time. Most of the field lands between 10 and 30 seconds.

Which image model should you use?

• For most things, Gemini Nano Banana 2.1 ($0.034). It is on the best-value line, edits people faithfully and takes under 20 seconds.
• For lots of images on a budget, Muse Image 1.0 ($0.01). It scores close to the leaders.
• For the best image regardless of price, GPT Image 2.5 Sunburst ($0.21). It tops both arenas, and you pay roughly 6x for the last 46 Elo.
• For edits of real people, Gemini Nano Banana 2.1, MAI-Image-2.6 or Grok Imagine Image 2.0. Avoid GPT Image 1 Mini.
• For swimwear, fashion or art with nudity, Gemini Nano Banana 2.1 was the only model that drew all of it.
• For instant drafts, Recraft V4.1 Flash ($0.007, about two seconds).