We gave every AI image model we could reach the same four prompts: a text-heavy recipe page, a labelled engine diagram, a hand-drawn map of Onomichi, and a barista pouring latte art. One shot each, no re-rolls.
Models are ordered by release date, newest at the top to oldest at the bottom, so scrolling down is a trip back in time. Watch the text in each image: it is the clearest sign of how fast this is moving. Click any image to open it full size. This is a living benchmark we re-run as new models ship. 😺
GPT Image 2
Released 17 April 2026 · ~60s per image · ~$0.19 · text: excellent
OpenAI’s flagship and, by date, the newest model here. The most detailed and illustrative of the bunch, and fully legible. The catch is speed and price: a minute or more per image, at nearly five times Gemini’s cost.
Gemini 3.1 Flash Image (Nano Banana 2)
Released February 2026 · ~12s per image · ~$0.04 · text: excellent
The value champion. Legible text, correct diagrams, lovely maps and believable photos, all for about four cents an image in twelve seconds. If you only try one, try this.
Gemini 3 Pro Image (Nano Banana Pro)
Released 20 November 2025 · ~24s per image · ~$0.12 · text: great
Powerful, but a reminder that “Pro” is not automatically better. On the recipe it duplicated a whole line, something the cheaper, faster 3.1 Flash that came after it got right.
Gemini 2.5 Flash Image (Nano Banana)
Released 26 August 2025 · ~9s per image · ~$0.039 · text: weak on dense text
The original Nano Banana that went viral last summer. Charming compositions, but lean in and the text starts to dissolve.
Imagen 4 Fast
Released 15 August 2025 · ~6s per image · ~$0.02 · text: weak on dense text
The oldest model we could still run, and where we were a year ago. Gorgeous layouts and solid photos, but any real text comes out as decorative nonsense.
The pattern
- Text broke through in 2026. A year ago a recipe page came back as handsome gibberish. Now it is printable.
- Hands are solved. The benchmark that defined “AI weirdness” is a non-issue, even on last year’s budget models.
- Flash beats Pro. The smaller, cheaper, more recent Flash models matched or beat their Pro siblings more than once.
- Gemini is cheaper and faster. A perfect recipe page took Gemini 3.1 Flash twelve seconds and four cents; GPT Image 2 took a minute or more at nearly five times the price.
We would have loved to go back further, all the way to DALL·E 2 and 3, to show the full arc. Sadly OpenAI has retired them from the API, so 15 August 2025 is as far back as a live test can reach today.
Try it yourself
Every image here was generated through a plain API call, and all of these models, GPT Image and Gemini alike, are available right inside WordPress through AI Engine. Generate from a prompt, drop the result into your Media Library, and switch providers without touching code. ✨
Last updated: June 2026. Release dates from each provider; OpenAI dates are exact from the API, the Gemini 3.1 Flash day was not published. One prompt, one generation per model, no re-rolls.
