The pricing of image generation by LLMs has always confused me. It starts with me not understanding how an image can be priced by text tokens. Now that OpenAI released GPT-image-2, I wanted to know if switching over from Nano Banana 2 would make sense for the storybook project I have running with images of my kids.
The first image I generated was promising. Really promising. So I asked ChatGPT to estimate what a full book spread would cost.
The answer was not what I hoped for. For one final 42 × 21 cm spread at 300 dpi, the budgeting number came back at about $0.65 to $0.75 in medium quality, and $2.50 to $2.65 in high quality. My book has 12 spreads and a cover. The high-quality number would not make financial sense for a single gift book.
So I had to think about concessions. Medium quality for a watercolor style could probably work. But I prefer the highest setting. I also do not fully trust the ChatGPT estimate. Maybe the real way to know is to run one actual spread and see what the bill says.
Nano Banana 2 would have been cheaper. The issue is that GPT-image-2 is visibly nicer, and if I can make it work, the storybook would be noticeably better. Not a little better. Clearly better.
First I needed the tool to support the new model. I wanted to keep the old one as a fallback, because I do not know yet which model will end up being the final choice. So I asked Codex to add a model selector and it got to work.
After I added my API key and confirmed everything was wired up, I generated a new image for the book. The result was better than expected. The watercolor style actually made me believe, for the first time, that this book is going to be a great gift. That gave me enough confidence to keep working on the project today.
Last step was sending the final result to the print company. In a few days I will see it in real life and know if the quality survives the paper. I will keep you posted.
Key Insight
I talk a lot about budget as the reason to hold back, but the real thing that unblocks me is seeing one good image. The spreadsheet does not convince me to keep going. A single watercolor image of my kids does.