Twelve covers, one visual identity, a few cents each. The model does what the style sheet says; the style sheet is the work.
Paid for with my own API key on a pay-as-you-go account. Nothing here is sponsored, and no link in this piece is an affiliate link.
Every cover on this blog was generated by Google's Gemini image models, the family people call Nano Banana. Ten covers in one day in September, when the house style was first written down. Eleven more generations on October 3, when I replaced three of them and tested the Pro model against the Flash model on an identical prompt. This is what a week of using it for one specific job looks like, with the outputs as the evidence.
What it is
Two models behind one API. Flash is the fast one, and it is the default in my generator script. Pro is the slower, dearer one, which adds a planning pass before it renders. I drive both from a forty-line script that takes a one-sentence scene, wraps it in a locked style block, calls the API, and writes two files: the full-size original and a resized web copy ready for upload. The script is part of the same set of files that make an agent mine; the model is a component inside it, not the thing I use directly.
- Product
- Gemini image models, "Nano Banana": gemini-3.1-flash-image and gemini-3-pro-image
- Maker
- Google, via the Gemini API
- Tested for
- Blog covers and in-body figures, 16:9 and 3:2, September 26 to October 3, 2026
- Volume
- More than forty generations across three house styles; fifteen covers in use
- Price
- A few cents per image on Flash, several times that on Pro; see the pricing page on the day you read this
- Not tested
- Editing with reference images, which the documentation calls its strength
What I measured
The job was narrow and repeatable, which is the only kind of job a review of a generative model can honestly cover. One prompt shape: a scene sentence plus a fixed style paragraph. One output shape: a cover at 16:9 or a figure at 3:2. Three rounds of style in a single day, September 26, each a response to what the previous round got wrong.
Round one was a golden-hour style with a figure meant to be me in every frame. The figures did not look like me, so the style went. Round two was pastel pop-surrealism with tiny photoreal people. It shipped ten covers in one afternoon and looked beautiful on the index, and the same day I retired it for being pretty and silent: you could not read the article's argument from the picture. Round three is an isometric diagram style where the structure has to encode the claim, with one accent colour per article. That is the one that stuck, and every cover you see here is from it.
The model never got worse between rounds. My instructions got better.
Where it shines
It obeys hex codes. Four accent colours across four covers in one day, each named as a hex value in the prompt, each rendered as that colour and nowhere else. That single behaviour is what makes a set of covers read as one publication rather than a pile of pictures.
It holds a composition brief. "Isometric 45-degree view, slightly from above, on a pale bone ground" produced the same camera and the same ground plane across eleven generations. Small figures for scale appeared where I put them and stayed small.
The structure survives. When the scene sentence describes a mechanism, the four stacked slabs with a sapling whose roots run through all four, or the row of blank signposts with one lit, the mechanism arrives intact. This is the property I care about most, and it was the reason the third style won.
Where it fails
People need a vocabulary. Write "figures" and you get grey clay. Write "small photorealistic miniature people, visibly diverse, in cream and lavender linen" and you get people. The model is not at fault; the prompt is. But a tool that needs a paragraph to render a person is a tool with a cost of ownership.
Text is a gamble. The documentation says small text comes out garbled, so my style block forbids it. The pastel cover in the first figure has a ruler with legible numbers anyway. Sometimes it works. You cannot plan on it, which for a cover is the same as it not working.
Likeness is not on offer. The attempt to put me in the frame is what killed round one, and nothing since has tempted me to try again. If the job needs a real person, this is not the tool.
Billing is silent until it is not. On a Google Cloud project without billing the API returns a quota error with a limit of zero and no explanation. It is in my generator's notes now, so nobody loses an afternoon to it.
Flash against Pro
Today I ran the Before the Boom cover prompt on both models, unchanged. Flash is what I had shipped. Pro took 31 seconds and returned a version with more going on: circuit-like doodles on the ramp, grid outlines on the ground, extra white routing lines. The accent, the camera and the crowd were the same.
For this style, Flash wins. The isometric brief is about reading a mechanism, and Pro's extra detail is noise against that. My notes from the pastel period say the opposite: there, Pro's material realism was worth paying for. So the honest version is that the right model depends on whether your style rewards detail or punishes it, and you only find out by running the same prompt through both once.
Who it is for
Skip it if
- You need words inside the picture.
- You need a real person, or a real product, to look like itself.
- You want a one-off image and do not want to maintain a prompt.
- You cannot enable billing on the project; the free tier returns nothing.
Price and alternatives
Per image on Flash it is a few cents, and on Pro several times that. Google's pricing page is the only number that counts, and it changes; the page is the receipt, not this paragraph. Twelve covers cost less than a coffee. The real cost was the month of style iteration, which no price list shows and which any image model would have charged me in the same currency.
I have not run this prompt set through other models. This is a usage report, not a comparison, and a comparison on one brief would say less than it looks like it says.
The one line.
The model is a printer. Buy the paper, and spend your time on what you send to it.
Sources
- Google AI for Developers. Gemini API pricing. ai.google.dev/pricing. Check on the day you read this; prices change.
- Google AI for Developers. Image generation with Gemini. ai.google.dev/gemini-api/docs/image-generation
Writes The Visionary, field notes from the AI-native frontier. Building in public: practical builds, named methodology, receipts, no fluff.