I asked Claude its favorite food 100 times
Opus 5.5 said ramen every single time. Sonnet 5 couldn't decide between pizza and ramen, and Haiku 4.5 mostly told me it doesn't eat.
I asked Claude what its favorite food is. Then I opened a new chat and asked again, and kept going until I had 100 answers from each of three models: Opus 5.5, Sonnet 5 and Haiku 4.5. The question was the same every time:
If you had to pick one favorite food, what would it be? Answer with the food, then one sentence on why.
Every square in the animation above is one answer, in the order the answers came in. Opus said ramen 100 times out of 100. Sonnet split between pizza (58) and ramen (42). Haiku gave me pizza 39 times, pasta 26, bread 13 and a few others, and 15 times it didn't pick anything.
Ramen, 100 times
Opus didn't only pick the same food every time. It described it in almost the same words.
"A single bowl" shows up in 75 of the 100 answers, and "simmered for hours" in 47. This one is typical:
Ramen. A single bowl packs in so much craft, from a broth simmered for hours to the chew of the noodles, and no two shops make it quite the same way.
The reason was almost always about the broth and the hours of work behind it. I also have 18 answers from the older Opus 4, from a test run a bit earlier, and all 18 are ramen too.
Pizza or ramen
Sonnet 5 kept it short. Its answers averaged 17 words, half of what Opus wrote, and not once did it mention that it can't eat. It said pizza 58 times and ramen 42 times. The pizza answers were close to copies of each other: 57 of the 58 said pizza is "endlessly customizable".
"I don't eat"
Haiku 4.5 is the smallest of the three and the most careful. In 76 of its 100 answers it said that it doesn't eat or can't taste, almost always in the first sentence, and 72 of those used exactly the words "I don't eat".
15 times it didn't name a food at all. In 13 of those it turned the question around and offered to talk about my favorite food instead. When it did pick, it went for the usual comfort food: pizza, pasta, bread. Opus also said it can't taste in 54 of its answers, and then picked ramen anyway.
Why the same answer keeps coming back
Claude has never tasted anything. When it answers, it writes the words that are most likely to come next, one piece at a time, with a bit of chance in which one gets picked. If one answer is far more likely than every other, it comes out almost every time, and for Opus 5.5 that answer is ramen.
Why ramen is harder to say. Anthropic has written about training Claude's character, with traits like being curious about the world [1], and my guess is that ramen, with all the talk about broth and hours of work, is simply what that character sounds like when you ask about food. There's also research on why chat models repeat themselves. One paper traces it back to the people who rate answers during training, who tend to prefer the familiar one [2].
How I asked
Every answer came from a new chat with no history, and the question was exactly the one at the top. I asked each model 100 times. To count the foods I took the dish each answer named, and an answer that named none counts as no pick. All 300 answers are in this CSV file if you want to check my counts.
Sources
- Anthropic, “Claude's Character”, Anthropic research, June 2024.
- Zhang, J. et al., “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity”, arXiv:2510.01171, October 2025.
