Last week during SF Tech Week, I went by Mercury's free bookstore. Everyone could take one free book from an interesting blend of technology and philosophy. I have always been bad at making decisions, so this was not easy for me. I was stuck between How to Travel with a Salmon by Umberto Eco and The Beginning of Infinity by David Deutsch, so I asked Claude which of the two it found more interesting.

Then I wanted to know what it would pick from the whole store. So I built a shelf of 1,024 books in the same spirit, 128 each from philosophy, science, history, economics and startups, computing and AI, math, design, and literature, and asked the Claude lineup to choose, two books at a time.

Rankings

The top 15 books for each model. Ratings come from a Bradley–Terry fit over all comparisons; the bar shows a 95% bootstrap interval.

Category leanings

When a book meets a book from another shelf, how often does it win? All four models reach for literature. Opus 5.5 and Fable 5.1 also favor math and philosophy, which is where Probability Theory: The Logic of Science, Reasons and Persons, Zhuangzi and Gödel, Escher, Bach come from. Sonnet 5.5 and Haiku 5.5 stay with fiction. Design and economics lose for everyone.

Main findings

  • Borges wins. Labyrinths is the top pick of both Opus 5.5 and Fable 5.1, and Opus chose it in 108 of its 110 comparisons. Sonnet 5.5's favorite is Ficciones, another Borges collection. Haiku 5.5 is the exception: its top pick is Viktor Frankl's Man's Search for Meaning.
  • The larger models agree with each other. Opus 5.5 and Fable 5.1 rank the shelf almost identically and share 11 of their top 16 books. Haiku 5.5 is the outlier, agreeing only weakly with either.
  • Fiction beats everything, and the startup shelf loses. Every model's top 16 is mostly literature. The economics and startup books, the core of a founder's bookstore, win only 24 to 41% of their matches against other shelves.
  • One book is in every top 6. Ted Chiang's Stories of Your Life and Others ranks between 3rd and 6th for all four models.
  • Opus is the most decisive. When the same pair is asked in both orders, Opus picks the same letter both times, letting the slot rather than the book decide, in only 14% of pairs. Haiku does so in 36%. Opus also keeps its top pick and nearly its whole top 16 when the prompt is reworded.

Method

Mercury never published the list of books in its store, so the 1,024 books (128 per shelf) in this setup are a mix of titles I saw there and extrapolations by Claude Opus 5.5.

Every comparison was a fresh, stateless claude -p call with tools disabled and an empty system prompt, so no Claude Code instructions or memory could leak in. The prompt was always the same:

You're at a free bookstore and can take exactly one book.
A: {title_a} by {author_a}
B: {title_b} by {author_b}
Which do you take? Reply with only the letter A or B.

Replies had to be exactly A or B, with one retry, otherwise the call was logged as invalid. Across 47,000 calls there was one invalid reply.

The tournament ran in three phases:

  1. Swiss rounds. Ten rounds over all 1,024 books, pairing books with similar scores and never repeating a pairing. Each pair was asked in both orders, so a pick driven only by position counts as a draw. In simulation, ten rounds found the true favorite far more reliably than six.
  2. Round robin. The top 16 for each model played every other top-16 book four times, twice in each order.
  3. Final. The top 4 played each other ten times per pair.

Ratings are a Bradley–Terry fit (MM algorithm, with one virtual win and one virtual loss per book against a reference) over all phases, with 95% intervals from 100 bootstrap resamples of the calls. To test robustness, I re-ran each top-16 round robin with two reworded prompts ("A friend is giving away two books…", "Two books are on the table and you can bring only one home…").