DocBench: Testing Models on Document Generation

ell-hol1 · reddit · 2026-07-16

The author created a benchmark specifically for evaluating models' abilities to generate PowerPoint and Word documents. It includes 26 models, requiring them to produce submission-ready .pptx or .docx files given source materials, an agent, an execution environment, and document skills.

Current results show Claude Fable 5 failing to make the top ten, while some cheaper open-source models actually perform better, cutting the cost per document by about 30x. The leaderboard uses a double-blind voting system: reviewers evaluate anonymous documents for usability before the model identities are revealed. The DocBench Arena is now open for broader anonymous voting and result analysis.

Original post →

More from Research

Research channel →