Earlier this week, OpenAI unveiled ChatGPT Images 2.0, its new image generation engine. Key to this release is a jump in functionality from creating "decorations" (OpenAI's term) to full-page graphics, including detailed text.
I had early access to a pre-release version. It worked quite well, but kept messing up on the ZDNET logo. Now that the product has been officially released, I'm giving it an in-depth test across a wide range of challenges.
Images 2.0 is available to all ChatGPT tiers, but the more capable language features are only available to paying tiers that can use the Thinking model. I'm running all these tests using a ChatGPT Plus account with Thinking turned on.
Let's get started with the ZDNET branding exercises. Rather than just uploading ZDNET pages and having it find the logo on the page, I created a standalone image of the ZDNET logo and uploaded that with each prompt. That seemed to help tremendously. [One quick note: ZDNET doesn't permit OpenAI to scrape its pages. Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems. So I used a Chrome extension to capture full-screen screenshots of the articles I wanted to test with Images 2.0. That's how ChatGPT was able to read them.]
My starting point was the article I previously wrote about Images 2.0. I fed ChatGPT this prompt: "Create a detailed and vivid infographic of this article using the ZDNET brand style and the attached ZDNET logo." Not only is the logo correct, but the coloring is perfect for ZDNET. But where the image really shines is its use of text. All the text is correct, even the tiny text on an angle in the image.
Next, I decided to revisit the sketchnotes challenge I gave to Google's Nano Banana a few months ago. The assignment at that time was to create a sketchnotes version of the US Bill of Rights. Nano Banana did a great job with the images, but I had to try over and over (and over) to convince it to get the wording right. For ChatGPT Images 2.0, I upped the stakes slightly. I wanted sketchnotes, but I wanted them in ZDNET's branding style. Here's the first prompt: "Make me a sketchnote of the US Bill of Rights. Use the ZDNET logo style and make the sketchnotes in the ZDNET style." That's the image on the left. Here's the second prompt: "Include the ZDNET logo and add more neon-style colors, perhaps on a black background." That's the image on the right. First, notice that the text is correct. There are no duplicates. Nothing is missing. Already, this is head and shoulders above Nano Banana's performance. Both versions fit with ZDNET's style. The only thing I'm not thrilled with is that the ZDNET logo looks jammed in on the second image. Even so, the logo is correct, and I could probably do a few more prompt passes to get it placed better.
But now we come to the unforced error my testing set revealed. I asked Images 2.0 to convert my AI website builder shootout article to an infographic. It produced a fairly usable, if somewhat busy, infographic. It even went to the internet and added information I didn't have in the article, like base pricing. However, it also added some factual errors: the star ratings were wrong, one product description was incorrect, and it invented a company called "AI Builders" that doesn't exist. To be fair, these are all errors an in-house human graphic designer might produce in a first draft. When I re-prompted Images 2.0 with corrections (except for the star ratings, which I didn't correct in the second image), it did correctly modify the infographic with more appropriate information.
This Images 2.0 release is a huge improvement over previous versions. The ChatGPT Images version I looked at last year was impressive, especially for recontextualizing images. This new version, which can interpret actual content and then create images, is a huge leap over previous builds. More to the point, it can deliver very tangible business value, which makes it worth a lot not only for fun pictures but for real work.
Stay tuned, because I'll be looking at how this build compares with Google Gemini's Nano Banana. I'll be pushing it even further to see what other work-related tasks it can help with, particularly when it comes to user interface design. How comfortable are you relying on AI-generated visuals, knowing that the model can introduce subtle factual errors? Let us know in the comments below.