Blogger casens: LLMs Can Now Read Novels, Refuting Marcus's 2029 Prediction

Blogger casens published an article on October 13, 2026, claiming that current large language models can clearly read books, refuting earlier predictions by AI critic Gary Marcus. In 2022, Marcus proposed a bet asserting that by 2029, AI would not be able to reliably read a novel and answer questions about plot, characters, conflicts, and motivations. casens argues this prediction has already been disproven.

To verify this claim, casens selected "Hollywood, Ending," a novel by John Green published on September 22, 2026, as the test material. Because the book was released after the training cutoff for most models, it served as a true "unseen text" test. casens converted the novel into pure Markdown format, removing all metadata, and ran Claude Opus 4.6 in a sandboxed environment that blocked all internet traffic except for API connections.

The results showed that Claude Opus 4.6 performed excellently on short-answer questions about the novel, producing no hallucinations and including all intended answers. On long-answer questions, the model's performance was rated by casens as equivalent to at least an "A" grade (95 out of 100) in a high school or university literature class. Despite minor discrepancies in details or differing interpretations of themes, the model successfully identified main characters, motivations, conflicts, and the overall plot without major omissions.

casens also cited other benchmarks to support his view, noting that QuALITY, Winogrande, and DROP—tests designed to measure reading comprehension and common-sense reasoning—were already saturated with high scores by GPT-4 and Claude Opus 3. He argued that the failure cases listed by Marcus, such as understanding pronoun references or inferring implicit information, were limitations of AI from 2019 and earlier, not characteristics of current frontier models.