Some friends brought over the game Cover Your Assets last weekend for a game night with the kids. At the end of the first hand, everybody started counting at once, out loud, as kids do.

So I thought, Let’s see how Claude does at scoring a hand. I took a picture, uploaded it to the app, and gave it a simple prompt: “Count the value of all of these cards, in dollars.”

Claude’s estimated total: “approximately $560,000.”

Cool. I have the ChatGPT app on my phone, too, so let’s try that with the same photo, same prompt. “So, the total value of the cards is $430,000,” says the second robot.

That’s interesting. I guess we better try more.

I used llm and a little Python to loop through 21 OpenAI and Anthropic models that all support JPEG attachments. All the code is on Github.

The results are all over the place, from $345,000 using gpt-4.1-nano to $545,000 using claude-sonnet-4-0. All of the individual results were in the repo, showing full responses and token usage.

I’ve been excited lately about the possibility of using AI and computer vision for reporting, and large language models can do amazing work in this space. But this was a good reminder that results need to be checked, and carefully.

The correct answer here is $465,000.