I asked ChatGPT to make an image of a completely full glass of wine. It did

OpenAI released the latest–and massively upgraded–version of ChatGPT’s image generation engine on Tuesday, and the internet was soon oohing and aahing, giddily asking the AI to make everything from memes in the style of South Park to images of Barbie dolls in the Oval Office.

But one feat of ChatGPT’s new GPT-4o image generation model left even jaded AI watchers in a state of hushed, slack-jawed awe.

Red wine, anyone?

Behold, ChatGPT can now—quite reliably—render an image of a glass of red wine filled to the very tippity-top.

Prompt: render an image of a wine glass filled to the very top with red wine

Ben Patterson/Foundry

Sounds like a simple task, right? Surprisingly, the “full glass of wine” test has stumped plenty of big-name AIs, including—until now, anyway, ChatGPT and its older DALL-E engine.

Here, for example, is Google’s Imogen 3 flubbing the test when using the same prompt:

Ben Patterson/Foundry

And Grok 3 doesn’t fare much better:

Ben Patterson/Foundry

Microsoft’s Copilot also took a stab:

Ben Patterson/Foundry

I even tried with Flux, one of the latest Stable Diffusion models, and got this:

Ben Patterson/Foundry

Whoops.

The “glass of wine” trick isn’t a formal benchmark of an AI’s image-rendering abilities; instead, it’s a casual test, like asking an LLM how many “r’s” are in the word “strawberry.” They tend to get it wrong, sometimes hilariously so.

Why is a completely full glass of wine such a challenge for image-generating AIs? The prevailing wisdom is that AI-powered models do best with images they’ve been trained on—and when it comes to pictures of red wine glasses, they’re typically filled about halfway, which is why a prompt for a “COMPLETELY full glass of wine, all the way to the brim” tends to get you a half-full glass.

Now, a really good AI image generator should (as one Redditor helpfully explained) be able to “extrapolate” the idea of a completely full glass of wine even if none exist in its training data. Either that, or someone at OpenAI just fed the new model dozens of pictures of filled-to-the-brim wine glasses.

Of course, there’s another acid test for AI image generators: an analog clock set to a specific time. Betcha ChatGPT and its new image generator can make short work of that one, right? Let’s see:

Prompt: render an image of a clock, with the hands showing 3:15