ChatGPT’s New Image Generator Is Better, but It’s Still No Nano Banana

Back in March, while I was testing and reviewing the first iteration of OpenAI’s image model, I gave the chatbot what I thought would be a simple test: Make me an infographic explaining the water cycle. That test ended up being something of a disaster. So when OpenAI announced its new 1.5 model and ChatGPT Images interface, I was ready to give it another go. Sadly, what I found was emblematic of the progress and persisting issues with generative media in 2025. Stick with me a minute and recall when you first learned about the water cycle. For me, it was in elementary school, and we had a rhyming song to help us remember how water is transformed throughout the environment, from water vapor to clouds to rain. I vividly remember looking at different images and videos in class that explained the different phases. So, for a company that says its chatbot’s GPT-5 is as smart as a Ph.D., why can’t its image models render a simple graphic of one of the most basic (or should we say, elementary) and well-documented geological processes? (Disclosure: Ziff Davis, CNET’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.) This is much improved for ChatGPT, but it’s still not perfect. The arrows aren’t facing the correct directions. Created by Katelyn Chedraoui using ChatGPT If I were to defend ChatGPT, I would point out that this newest infographic is much better than previous versions. The text is finally legible, a huge step up from even a month ago. It included all the phases and its sources, as requested. But the arrows are still wrong, with precipitation’s vertical-facing arrow giving the impression that rain can fly upward from the ground. I asked it three more times after its initial fail, and the results were virtually identical. With the original OpenAI image model, this took me more than 20 prompts and 2 hours to create. There are still errors. Created by Katelyn Chedraoui using ChatGPT I am positive that nobody using the new ChatGPT Images feature is as invested as I am in its ability to correctly render the water cycle. Perhaps I’m being overly picky. But the problem highlights an ongoing issue with OpenAI: It may have one of the most widely used chatbots, but its creative services leave much to be desired. That didn’t used to be a big problem — until Google changed the game this summer. This ChatGPT-made holiday card preserves my friends’ and my features from the original photo (left) pretty well, but our skin still has that shiny AI look. Created by Katelyn Chedraoui using ChatGPT

OpenAI’s ChatGPT Images vs. Google’s Nano Banana

OpenAI and Google have been duking it out throughout 2025 in the race for the most advanced AI. Google, for the most part, has taken the lead, but OpenAI has never been far behind. When Google dropped industry-leading video and image models, like Veo 3 and Nano Banana, OpenAI would follow up a few weeks later with Sora and now Image 1.5. This image model in particular, along with GPT 5.2, is likely the result of an internal “code red” OpenAI reportedly issued out of concern for the technical prowess Google displayed with its recent family of models in Gemini 3. In other words, the top two tech companies, though not traditionally known for their creative history, have been investing heavily in creative AI. It has led to massive technical advancement — and a lot of concerns. ChatGPT Images (right) is improved over previous versions, but it isn’t fooling anyone the way Gemini’s AI images are (left). Created by Jon Reed using Gemini and ChatGPT Google’s Nano Banana Pro is our recently crowned best AI-image generator of 2025, and it was one of the most impressive, awe-inspiring AI programs I’ve ever used. It wasn’t perfect, and certainly left me with queasy feelings, but it did the job it should’ve been able to do. By comparison, ChatGPT Images with the new 1.5 model marked a huge leap forward for the company but didn’t wow me in the same way. Gemini (left) was able to make a more realistic version of my family in the Appalachian State University’s Kidd Brewer Stadium, while ChatGPT’s is clearly fake with its over-the-top snow. Created by Katelyn Chedraoui using Gemini and ChatGPT Gemini generates images quicker than ChatGPT, and those Gemini images tended to be better in my testing. I also hit my generation limit with the free version of ChatGPT quicker than I did with the free version of Gemini, which was annoying. One area where ChatGPT does well is with less realistic, more creative work. OpenAI’s AI images became recognizable for the way they leaned into cartoon-like styles and a faint yellow tint on some images. The yellow tint is gone now, but the cartoon style remains a place where ChatGPT excels. It certainly has a lot of experience with it, since its first model went viral with people making Studio Ghibli-esque versions of themselves. In my short time using the new model, I could see where OpenAI is going. It’s already shown big improvements, especially with text and realistic photography. But I can’t help but wonder if this 1.5 model is a half measure to prove it can compete with Google and that even better abilities await in the full second-generation model. We’ll have to wait and see. For now, my advice remains the same: ChatGPT Images is great for ChatGPT users, but Nano Banana Pro is truly exceptional.