ChatGPT's updated image model tested against Google's Gemini Nano in nine side-by-side business use cases
Nate B Jones of AI News & Strategy Daily runs a head-to-head image generation comparison between ChatGPT 5.2's new update and Google's Gemini Nano Banana Pro across nine business-relevant tests.
Summary
Nate B Jones frames this comparison as part of ChatGPT's ongoing "code red" response to Google's Gemini releases, with OpenAI claiming their new image update is up to four times faster and delivers superior editing capabilities. Jones tests both models across nine challenges — including celebrity image editing, children's alphabet charts, funnel diagrams, fictional maps, advertisements, ARR revenue bridges, Venn diagrams, opportunity solution trees, and pie chart editing — and finds that Gemini Nano Banana Pro outperforms ChatGPT 5.2's updated version in nearly every category. A key structural finding is that ChatGPT appears to generate complex diagrams by writing code and then photographing the result, which leads to cut-offs, misalignments, and reasoning failures, while Gemini Nano Banana Pro appears to reason through image generation more directly. Jones concludes that Gemini Nano Banana Pro is currently the only image model he would trust for serious business work, and advises readers to run their own tests rather than rely on benchmarks.
Key Takeaways
FULL TRANSCRIPT
Context and methodology
Nate B Jones: ChatGPT continues a code red response to Google. For context, they've been in this code red mode for a while since Google launched Gemini. ChatGPT 5.2 was the initial response to that, and now they're continuing with a new images release that is of course aimed at Gemini Nano Banana Pro. ChatGPT is claiming faster image generation — up to four times faster — and they are obviously saying that theirs is "better" and that it's going to be able to deliver more compelling edit capabilities. I put all of that to the test. I went through and did a side-by-side comparison across nine different challenges with business-relevant implications, and I have to say Gemini Nano Banana Pro wiped the floor with ChatGPT 5.2, even the new updated version. I will show you the slides in a minute with side-by-side image comparisons, and you will see for each of the nine why Gemini Nano Banana Pro did a better job.
Before we get into that, just a couple of high-level observations. Number one, there is a different method that ChatGPT is using to generate these images, and I don't think it works well for them. This is particularly true for images that require a lot of logical thinking by the model. So if you ask it to develop a diagram that would be appropriate for a PowerPoint slide, Gemini Nano Banana Pro appears to use reasoning baked into the image generation process itself, and if it fails, you see a badly conducted set of reasoning with incorrect labels or something like that — and it actually doesn't fail very often. On the other hand, with ChatGPT, what you see is code. If it fails, you see code. It is literally writing the code for the diagram and then trying to photograph the results and bring that to you. That has concrete consequences, and you'll see issues with lining up the diagrams in a way that the model can photograph. The model clearly doesn't understand quite what it's doing. There's not an internal reasoning check.
It looks like ChatGPT tried to compensate for this by including a self-edit loop in this launch. When I did a children's alphabet test — where you have an A for aardvark, an animal for each letter, going all the way through A to Z — ChatGPT tried to catch itself and edit itself. It got into a 20-minute edit loop. It produced about a dozen images. And at the end, the resulting quality was still not any better than the initial image. I like the idea of checking and rechecking the work, but I'm not seeing actual quality gains that would justify that kind of time.
Despite the claim that this is a very fast image generator, I found in practice that Gemini Nano Banana Pro generated the images I'm about to show you much, much faster and with a lot less drama, a lot less thinking, a lot less reasoning. They just got it done and generated an image.
I'm not going to tell you Gemini Nano Banana Pro is perfect — you're going to see a few issues as we go through this slide deck. But overall, there is a tipping point where an image model becomes useful for creating a useful PowerPoint slide, and I have several examples in the deck here. Gemini Nano Banana Pro has hit that tipping point, and ChatGPT 5.2 isn't there. No other image model is there today. No other image model is as good as Gemini Nano Banana Pro today. So with that, let's hop in and see a comparison across nine different slides, side by side.
Test 1: Celebrity image editing — Keira Knightley
Okay, here we have a dual test. I wanted the model to take a celebrity and be able to repurpose her into a different location. This is an image edit test. I used Keira Knightley because her image is going to be widely available in training data, and I wanted to see if the model could adequately present her in an unusual situation — in this case, teaching how LLMs work. This allows me to test whether the model can show a diagram within the image, whether it can handle the perspective shift, and of course whether it can handle representing a celebrity's likeness correctly.
You might think: why are we worrying about celebrities? This is relevant because if you include an image of yourself, you want to know if it's going to look like you. That was really the test. I did not call her Keira by name because I didn't want to run into any copyright issues. All I did was give it a blurry picture of Keira Knightley from Pirates of the Caribbean and said, "Please have her teach how LLMs work" — to both models.
What you get on the right from ChatGPT is not really a correct image of Keira Knightley. You get an overall nice, colorful, very high-level view of how LLMs work. With ChatGPT 5.2's approach, it's clear that Gemini Nano Banana Pro knows Keira Knightley. That is a photographically correct image of her. She's even in costume — this was not a visible costume in the source image, so it decided to put her in that costume and clearly knew the movie I was referencing. And then it has a much more detailed diagram of how LLMs work, although it's not quite as visually appealing.
Test 2: Children's alphabet chart
Let's go to the children's alphabet. On the left you see Gemini Nano Banana Pro, on the right ChatGPT. Both models failed, but they failed in interesting ways. What you'll see is that Gemini Nano Banana Pro needed this to be a complete box, and so it had Fox and Gorilla, and it had Fox and Goat — F, N, G, F, N, G. Individually, these are correct in their cells, but you don't need to repeat those letters. It did take some coaching. I will say in both cases I had to ask for edits, because the initial versions messed up the X — Gemini Nano Banana Pro presented X-ray, so we had some issues there.
I would say the ability to get to a final result was a little better from Gemini Nano Banana Pro, but not perfect. And ChatGPT kind of fell apart here — Zebra rendered indifferently, then some form of W at the end, then an X way down there. We just didn't get where we needed to go, and this is after multiple edits. So Gemini Nano Banana Pro again did a better job, although neither model was perfect.
Test 3: Funnel diagram slide
Let's go to the professional funnel diagram slide. This is quite a detailed slide. If you look, the text is all readable on the Gemini side. I can read "completion down 1.2 percentage points week over week, drop-off on password and SSO step" — that is a perfectly correct assessment of a leak in the funnel. What you see on the ChatGPT side is somewhat less text and a sort of weird funnel illustration. This does not look like the biggest leak in the funnel, even if mathematically 57 is the biggest drop-off from 820.
The thing I really want to call out from a quality perspective is that Google has taken the time to draw the entire sequence of graph charts correctly. This is graphed in such a way that it believably goes up and down point to point across dozens of data points. The ChatGPT version is just a very light overall version that clearly isn't designed to be a fully functional graph. From a level of detail perspective, Gemini Nano Banana Pro wins here. This is a case where the ChatGPT output is going to look good initially, and then you're going to dive in and say, "Well, it's not quite right" — and not quite right is not going to work with an image, because you would have to just regenerate it from scratch.
Test 4: Fictional map — P.G. Wodehouse's England
Let's look at fictional maps. This measures the LLM's ability to generate spatial relationships and understand how story structures work. I chose P.G. Wodehouse's England because it's a very well-known corpus of books that the models have read, but it's not often mapped. It's not like Lord of the Rings, where there's an obvious map to reference in the training data.
In this case, I think Gemini Nano Banana Pro knocked it out of the park. All of these funny-sounding names are actually in P.G. Wodehouse's novels. Lord Emsworth is associated with Blandings Castle in the novels, and Bertie Wooster is associated with Brinkley Court, as is Aunt Dahlia. So it got it right — it got the characters correct and associated them with the correct locations in the novels.
On the other hand, ChatGPT really struggled. It initially named and generated a bunch of points on a map and tried to generate a photograph of a paper map, but if you zoom in, it is so blurry and tiny that you can't read it even zoomed in. There's nothing really usable about it. It's just a nice visual concept of a map. And that's kind of the whole game right there — you have to be able to generate a map and actually make it readable. There may be a comprehension issue here as well. This may be a situation where ChatGPT took the ask very literally and wanted to list out a bunch of place names, whereas Gemini Nano Banana Pro was able to synthesize more effectively across the ask.
Test 5: Advertisement layout
Advertisements — this is perhaps more business-relevant. Gemini Nano Banana Pro and ChatGPT both did pretty well here. The choice of aspect ratio and layout was left to the models. I think the overall layout worked better on Gemini Nano Banana Pro — that nice four-badge arrangement all the way across over the car looks really good, and the car is centered nicely. The ChatGPT version is still a fine ad. I don't think there's a huge issue, just a small one where "safe pickup and drop-off" wasn't handled correctly because it had to be dropped underneath the three badges. But overall, not too bad on either count.
Test 6: ARR revenue bridge
ARR revenue is a real problem. Gemini Nano Banana Pro correctly built a revenue bridge. A revenue bridge is very simply your starting ARR, with green upward marks for all of the additional ARR you get from new and expansion revenue, and then red for contraction and churn, and then your ending ARR. That's just how it is — it's a very defined chart style.
In this case, you'll see that example of ChatGPT trying to code this, because it could not photograph what I'm sure it coded, which was an ARR bridge. It cut it off at "RR" and also cut off the notes section. So that's not going to work. You cannot recover that — I checked, and the image is the image. This is just lost. And worst of all, the 4.2 should not be going down to 4.5. It should not have placed upward gains in revenue as declines in revenue. So it just misunderstood the assignment, and this is absolutely not usable.
Test 7: Venn diagram
The Venn diagram is another case where Gemini Nano Banana Pro just won straight up. I deliberately gave a challenging prompt that would not have been in the training data. I said, "Please create a Venn diagram of Taylor Swift, product managers, and the Army Corps of Engineers, and make it funny."
I got a fairly usable Venn diagram from Gemini Nano Banana Pro. A little bit wordy, but you can see what it's trying to do. It talks about coordinating massive high-stakes operations for all three. For Taylor Swift and the Army Corps, they're designing massive, structurally sound stages and infrastructure, and managing leaks — which was a nice funny touch.
The ChatGPT version just falls apart. There are no visuals to it. The model is trying to understand what it's supposed to do, but it wasn't able to make it funny, it wasn't able to draw it, and ultimately this is not something that would be usable. Again, you notice the cut-off issue — that's not me taking a bad screenshot, that's how it was produced.
Test 8: Opportunity solution tree
Let's try an opportunity solution tree. In this case, you get a full diagram from Gemini Nano Banana Pro — full text all the way through, very consistently styled, representing a usable solution tree for onboarding and activation. On the ChatGPT side, you get less detail, fewer options, and cut-offs that would make this unusable. It's almost as if it coded it again and just cut off what it was able to see from a coded series of boxes. This would not be usable on a slide because no one's going to accept the dot-dot-dot-dot-dot-dot, and Gemini Nano Banana Pro understands that and just writes it out.
Test 9: Pie chart editing
Let's try editing — that's one of the things that was highlighted as a strength of ChatGPT, that it can edit well. I took a diagram showing juice blend composition and simply said, "Please add 20% blueberries and make it correct." Gemini Nano Banana Pro was able to do that. Orange plus lemon plus grapefruit now equals 80%, and the blueberries equal 20%. This is a believable-looking pie chart. I believe Gemini Nano Banana Pro even got the 20% pie slice a little bit wider than the grapefruit at 15% and narrower than the lemon at 25%, so I think it did a fine job.
On the other hand, ChatGPT couldn't do it. It correctly added up — 24 plus 16 plus 40 is 80, and then blueberries are 20, so the math was fine — but it could not draw the pie chart. It just had blueberries spilling out everywhere. The grapefruit isn't correctly framed. This just doesn't work. And one of the smaller details I noticed is that Gemini Nano Banana Pro correctly put a little blueberry-purple tinge into the drink, and ChatGPT did not figure that out.
Overall verdict
So overall, my takeaway here is pretty simple: do not listen to the benchmarks. Do your own tests. And for now, Gemini Nano Banana Pro remains the only image model that I would trust for serious business work.