NextChapterBench Tests Whether Models Can Continue a Novel

Field notes

The benchmark supplies each model with a public-domain novel chapter plus a deliberately divergent brief, then compares the resulting continuations across compliance and craft. Its standings preserve operational detail often omitted from model comparisons: reasoning-effort setting, CLI route, output word count, per-run cost, and uncertainty margins. The visible table covers 13 models, 10 items, and three LLM judges, with links leading to model detail pages and the underlying chapters.