Questionnaire Question Extraction Benchmarks
2 min read
Measured results for Questionnaire question extraction. These pages are benchmarks: what was tested, on which model, and what to expect. Extraction turns an uploaded questionnaire (PDF, DOCX or XLSX) into the list of questions the Agentic Table answers, with each question's guidance and section context.
What users get. On the 18 standard documents, the best models capture 91% of questions (GPT-6.1 Terra, GPT-5.6 Terra and GPT-6 Sol), and the differences between them are not significant. With 18 documents the tests separate only large differences: even Claude Sonnet 5.5, at 58%, is not significantly different from GPT-6.1 Terra. GPT-4o (no reasoning, production default) captures 31%. 14 of the 20 models complete all 18 standard documents. Claude Haiku 5.5, Claude Sonnet 5 and Claude Sonnet 5.5 do not complete 4, 4 and 2 of them, because their replies are cut off at the platform's 10,000-token output limit; Gemini 4 Argon does not complete 9, because it returns an empty question list on them; GPT-5.1 does not complete 1, because its reply did not match the extraction format; Kimi K3 does not complete 1, because its requests on that document did not return. Claude models cannot be used for extraction in the product as it is, because Anthropic rejects the extraction schema; they were scored with the schema in the prompt. Five documents are larger than the standard route reads in one request; they are left out of the comparison because their result reflects document size, not the model, and the report's notes explain how each was handled.
Extraction models compared on the 18 standard documents, 1,495 questions (October–November 2026). Every model runs at low reasoning effort; GPT-4o, the production default, is not a reasoning model.
Recall on the 18 standard documents against mean list-price cost per completed document (log scale). Models with complete cost evidence on the documents they completed only (GPT-5.6 Terra, GPT-6 Luna, GPT-5 Mini, GPT-5.2, GPT-5.4 and Gemini 4 Argon left out). Cheapest at each level of recall, among the models that complete every standard document: GPT-5.6 Luna and GPT-6.1 Terra. GPT-4o's cost comes from its October 2026 run.
Recall on the 18 standard documents against median extraction time per completed document (log scale). The median, because one slow document can dominate a model's mean; the report gives both. GPT-4o's time comes from its October 2026 run.
Questions captured on the 18 standard documents, with 95% intervals. A document a model does not complete counts as zero, which pulls Gemini 4 Argon, Claude Haiku 5.5, Claude Sonnet 5, Claude Sonnet 5.5, GPT-5.1 and Kimi K3 down.